This page intentionally left blank
Inverse Modeling of the Ocean and Atmosphere Inverse Modeling of the Ocean and Atm...
17 downloads
369 Views
3MB Size
Report
This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!
Report copyright / DMCA form
This page intentionally left blank
Inverse Modeling of the Ocean and Atmosphere Inverse Modeling of the Ocean and Atmosphere is a graduatelevel textbook for students of oceanography and meteorology, and anyone interested in combining computer models and observations of the hydrosphere or solid earth. The scientiﬁc emphasis is on the formal testing of models, formulated as rigorous hypotheses about the errors in all the information: dynamics, initial conditions, boundary conditions and data. The products of successful inversions include fourdimensional multivariate analyses or maps of, for example, ocean circulation ﬁelds such as temperature, pressure and currents; analyses of residuals in the dynamics, inputs and data; error statistics for all the analyses, and assessments of the instrument arrays or observing systems. A stepbystep development of maximallyefﬁcient inversion algorithms, using ideal models, is complemented by computer codes and comprehensive details for realistic models. Variational tools and statistical concepts are concisely introduced, and applications to contemporary research models, together with elaborate observing systems, are examined in detail. The book offers a review of the various alternative approaches, and further advanced research topics are discussed. Derived from the author’s lecture notes, this book constitutes an ideal course companion for advanced undergraduates and graduate students, as well as being a valuable reference source for researchers and managers in the theoretical earth sciences, civil engineering, and applied mathematics. Tutors are also directed towards the author’s ftp site where they may download complementary overheads for class teaching. was awarded a Ph.D. in applied mathematics from Harvard University in 1971. He subsequently continued his research as a National Research Council Fellow at the University of Toronto, and as a Queen’s Fellow in Marine Science at Monash University (Melbourne, Australia). Following eight years as a lecturer and senior lecturer in the Department of Mathematics at Monash University, he became a research scientist at the Institute of Ocean Sciences, Sidney, B.C., Canada. He has been a professor at the College of Oceanic and Atmospheric Sciences at Oregon State University since 1987, where his research interests include ocean data assimilation, turbulence theory, and regional modeling. Professor Bennett has won refereeing awards from the Journal of Physical Oceanography (1986) and the Journal of Geophysical Research (1995), and is also the author of Inverse Methods in Physical Oceanography (Cambridge University Press, 1992). ANDREW BENNETT
Inverse Modeling of the Ocean and Atmosphere Andrew F. Bennett College of Oceanic and Atmospheric Sciences Oregon State University
The Pitt Building, Trumpington Street, Cambridge, United Kingdom The Edinburgh Building, Cambridge CB2 2RU, UK 40 West 20th Street, New York, NY 100114211, USA 477 Williamstown Road, Port Melbourne, VIC 3207, Australia Ruiz de Alarcón 13, 28014 Madrid, Spain Dock House, The Waterfront, Cape Town 8001, South Africa http://www.cambridge.org © Andrew F. Bennett 2004 First published in printed format 2002 ISBN 051103024X eBook (Adobe Reader) ISBN 0521813735 hardback
to Elaine, Luke and Antonia
Contents
Preface Acknowledgements
xiii xxi
Preamble P.1 Linear regression in marine biology P.2 Data assimilation checklist 1
1 1 4
Variational assimilation 1.1 Forward models 1.1.1 Wellposed problems 1.1.2 A “toy” example 1.1.3 Uniqueness of solutions 1.1.4 Explicit solutions: Green’s functions 1.2 Inverse models 1.2.1 Overdetermined problems 1.2.2 Toy ocean data 1.2.3 Failure of the forward solution 1.2.4 Leastsquares ﬁtting: the penalty functional 1.2.5 The calculus of variations: the Euler–Lagrange equations 1.3 Solving the Euler–Lagrange equations using representers 1.3.1 Leastsquares ﬁtting by explicit solution of extremal conditions
7 7 7 8 9 10 12 12 12 13 14 15 18 18
vii
Contents
viii
1.3.2 The Euler–Lagrange equations are a twopoint boundary value problem in time 1.3.3 Representer functions: the explicit solution and the reproducing kernel 1.4 Some limiting choices of weights: “weak” and “strong” constraints 1.4.1 Diagonal data weight matrices, for simplicity 1.4.2 Perfect data 1.4.3 Worthless data 1.4.4 Rescaling the penalty functional 1.4.5 Perfect dynamics: Lagrange multipliers for strong constraints 1.5 Regularity of the inverse estimate 1.5.1 Physical realizability 1.5.2 Regularity of the Green’s functions and the adjoint representer functions 1.5.3 Nondiagonal weighting: kernel inverses of weights 1.5.4 The inverse weights smooth the residuals 2
Interpretation 2.1 Geometrical interpretation 2.1.1 Alternatives to the calculus of variations 2.1.2 Inner products 2.1.3 Linear functionals and their representers; unobservables 2.1.4 Geometric minimization with representers 2.1.5 Equivalence of variational and geometric minimization: the data space 2.2 Statistical interpretation: the relationship to “optimal interpolation” 2.2.1 Random errors 2.2.2 Null hypotheses 2.2.3 The reproducing kernel is a covariance 2.2.4 “Optimal Interpolation”, or best linear unbiased estimation; equivalence of generalized inversion and OI 2.3 The reduced penalty functional 2.3.1 Inversion as hypothesis testing 2.3.2 Explicit expression for the reduced penalty functional 2.3.3 Statistics of the reduced penalty: χ 2 testing 2.4 General measurement 2.4.1 Point measurements 2.4.2 Measurement functionals 2.4.3 Representers for linear measurement functionals 2.5 Array modes 2.5.1 Stable combinations of representers
19 19 23 23 24 24 24 25 26 26 26 27 29 31 32 32 32 33 34 35 37 37 37 38 38 41 41 42 43 45 45 45 46 48 48
Contents
ix
2.5.2 Spectral decomposition, rotated representers 2.5.3 Statistical stability, clipping the spectrum 2.6 Smoothing norms, covariances and convolutions 2.6.1 Interpolation theory 2.6.2 Leastsquares smoothing of data; penalties for roughness 2.6.3 Equivalent covariances 2.6.4 Embedding theorems 2.6.5 Combining hypotheses: harmonic means of covariances
48 49 51 51 52 53 55 56
3
Implementation 3.1 Accelerating the representer calculation 3.1.1 So many representers . . . 3.1.2 Openloop maneuvering: a time chart 3.1.3 Task decomposition in parallel 3.1.4 Indirect representer algorithm; an iterative time chart 3.1.5 Preconditioners 3.1.6 Fast convolutions 3.2 Posterior errors 3.2.1 Strategy 3.2.2 Restatement of the “toy” inverse problem 3.2.3 Representers and posterior covariances 3.2.4 Sample estimation 3.2.5 Memoryefﬁcient sampling algorithm 3.3 Nonlinear and nonsmooth estimation 3.3.1 Double, double, toil and trouble 3.3.2 Nonlinear, smooth dynamics; leastsquares 3.3.3 Iteration schemes 3.3.4 Real dynamics: pitfalls of iterating 3.3.5 Dynamical linearization is not statistical linearization 3.3.6 Linear, smooth dynamics; nonleastsquares 3.3.7 Nonsmooth dynamics, smooth estimator 3.3.8 Nonsmooth estimator
58 59 59 59 61 61 62 64 66 66 67 69 71 73 74 74 75 76 78 81 82 83 84
4
The varieties of linear and nonlinear estimation 4.1 State space searches 4.1.1 Gradients 4.1.2 Discrete penalty functional for a ﬁnite difference model; the gradient 4.1.3 The gradient from the adjoint operator 4.1.4 “The” variational adjoint method for strong dynamics 4.1.5 Preconditioning: the Hessian 4.1.6 Continuous adjoints or discrete adjoints?
86 87 87 87 89 90 92 93
Contents
x
5
4.2 The sweep algorithm, sequential estimation and the Kalman ﬁlter 4.2.1 More trickery from control theory 4.2.2 The sweep algorithm yields the Kalman ﬁlter 4.3 The Kalman ﬁlter: statistical theory 4.3.1 Linear regression 4.3.2 Random errors: ﬁrst and second moments 4.3.3 Best linear unbiased estimate: before data arrive 4.3.4 Best linear unbiased estimate: after data have arrived 4.3.5 Strange asymptotics 4.3.6 “Colored noise”: the augmented Kalman ﬁlter 4.3.7 Economies 4.4 Maximum likelihood, Bayesian estimation, importance sampling and simulated annealing 4.4.1 NonGaussian variability 4.4.2 Maximum likelihood 4.4.3 Bayesian estimation 4.4.4 Importance sampling 4.4.5 Substituting algorithms 4.4.6 Multivariate importance sampling 4.4.7 Simulated annealing
94 94 95 98 98 98 99 100 102 104 104
The ocean and the atmosphere 5.1 The Primitive Equations and the quasigeostrophic equations 5.1.1 Geophysical ﬂuid dynamics is nonlinear 5.1.2 Isobaric coordinates 5.1.3 Hydrostatic balance, conservation of mass 5.1.4 Pressure gradients 5.1.5 Conservation of momentum 5.1.6 Conservation of scalars 5.1.7 Quasigeostrophy 5.2 Ocean tides 5.2.1 Altimetry 5.2.2 Lunar tides 5.2.3 Laplace Tidal Equations 5.2.4 Tidal data 5.2.5 The state vector, the crossover measurement functional and the penalty functional 5.2.6 Choosing weights: scale analysis of dynamical errors 5.2.7 The formalities of minimization 5.2.8 Constituent dependencies 5.2.9 Global tidal estimates
117 118 118 118 119 120 121 121 122 124 124 124 125 127
105 105 105 111 112 113 114 115
130 131 134 135 135
Contents
6
xi
5.3 Tropical cyclones (1). Quasigeostrophy; track predictions 5.3.1 Generalized inversion of a quasigeostrophic model 5.3.2 Weak vorticity equation, a penalty functional, the Euler–Lagrange equations 5.3.3 Iteration schemes; linear Euler–Lagrange equations 5.3.4 What we can learn from formulating a quasigeostrophic inverse problem 5.3.5 Geopotential and velocity as streamfunction data: errors of interpretation 5.3.6 Errors in quasigeostrophic dynamics: divergent ﬂow 5.3.7 Errors in quasigeostrophic dynamics: subgridscale ﬂow, second randomization 5.3.8 Implementation; ﬂow charts 5.3.9 Track prediction 5.4 Tropical cyclones (2). Primitive Equations, intensity prediction, array assessment 5.5 ENSO: testing intermediate coupled models 5.6 Sampler of oceanic and atmospheric data assimilation 5.6.1 3DVAR for NWP and ocean climate models 5.6.2 4DVAR for NWP and ocean climate models 5.6.3 Correlated errors 5.6.4 Parameter estimation 5.6.5 Monte Carlo smoothing and ﬁltering 5.6.6 Documentation
138 138
Illposed forecasting problems 6.1 The theory of Oliger and Sundstr¨om 6.2 Open boundary conditions for the linear shallowwater equations 6.3 Advection: subcritical and supercritical, inﬂow and outﬂow 6.4 The linearized Primitive Equations in isopycnal coordinates: expansion into internal modes; illposed forward models with open boundaries 6.5 Resolving the illposedness by generalized inversion 6.6 State space optimization 6.7 Wellposedness in comoving domains
172 172 173 175
References
186
Appendix A Computing exercises A.1 Forward model A.1.1 Preamble
195 195 195
138 141 142 142 143 144 146 148 150 156 165 165 167 168 168 169 170
177 180 182 183
xii
Contents
A.1.2 Model A.1.3 Initial conditions A.1.4 Boundary conditions A.1.5 Model forcings A.1.6 Model parameters A.1.7 Numerical model A.1.8 Numerical model parameters A.1.9 Numerical code and output A.1.10 To generate a plot A.2 Variational data assimilation A.2.1 Preamble A.2.2 Penalty functional A.2.3 Euler–Lagrange equations A.2.4 Representer solution A.2.5 Solutions for A.2.2–A.2.4 A.3 Discrete formulation A.3.1 Preamble A.3.2 Penalty functional A.3.3 Weighted residuals A.4 Representer calculation A.4.1 Preamble A.4.2 Representer vector A.4.3 Representer matrix A.4.4 Extremum A.4.5 Weights A.5 More representer calculations A.5.1 Pseudocode, preconditioned conjugate gradient solver A.5.2 Convolutions for covarying errors
196 196 196 197 197 197 199 199 199 199 199 200 200 200 201 205 205 205 206 209 209 209 209 209 209 210 210 211
Appendix B Euler–Lagrange equations for a numerical weather prediction model B.1 Symbols B.2 Primitive Equations and penalty functional B.3 Euler–Lagrange equations B.4 Linearized Primitive Equations B.5 Linearized Euler–Lagrange equations B.6 Representer equations B.7 Representer adjoint equations
212 213 214 214 216 217 219 220
Author index Subject index
221 225
Preface
Inverse modeling has many applications in oceanography and meteorology. Charts or “analyses” of temperature, pressure, currents, winds and the like are needed for operations and research. The analyses should be based on all our knowledge of the ocean or atmosphere, including both timely observations and the general principles of geophysical ﬂuid dynamics. Analyses may be needed for ﬂow ﬁelds that have not been observed, but which are dynamically coupled to observed ﬁelds. The data must therefore contribute not only to the analyses of observed ﬁelds, but also to the inference of corrections to the dynamical inhomogeneities which determine the coupled ﬁelds. These inhomogeneities or inputs are: the forcing, initial values and boundary values, all of which are themselves the products of imperfect interpolations. In addition to input errors, the dynamics will inevitably contain errors owing to misrepresentations of phenomena that cannot be resolved computationally; the data are therefore also required to improve the dynamics by adjusting the empirical coefﬁcients in the parameterizations of the unresolved phenomena. Conversely, the model dynamics must have some credibility, and should be allowed to inﬂuence assessments of the effectiveness of observing systems. Finally, and perhaps most compelling of all, geophysical ﬂuid dynamical models need to be formulated and tested as formal scientiﬁc hypotheses, so that the development of increasingly realistic models may proceed in an orderly and objective fashion. All of these needs can be met by inverse modeling. The purpose of this book is to introduce recent developments in inverse modeling to oceanographers and meteorologists, and to anyone else who needs to combine data and dynamics. What, then, is inverse modeling and why is it so named? A conventional modeler formulates and manipulates a set of mathematical elements. For an ocean model, the set includes at least the following:
xiii
xiv
Preface
(i) a domain in fourdimensional space, representing an ocean region and a time interval of interest; (ii) a system of inhomogeneous partial differential equations expressing the phenomenological dynamics of the circulation (there will be inhomogeneities owing to internal ﬁelds which force the dynamics beneath the ocean surface, and the equations will include empirical parameters representing unresolved phenomena); (iii) initial conditions for the equations, representing the ocean circulation or state at some time; and (iv) boundary conditions which may be inhomogeneous owing either to forcing of the ocean at the ocean surface, or to ﬁelds of ﬂow and thermodynamic conditions imposed at lateral open boundaries. There will be subtle yet profoundly important differences for, say, atmospheric models; consider the character of boundary conditions, for example. However, ocean models will be invoked henceforth as the default choice for concise discussion. From a mathematical perspective, the partial differential operators, initial operators and boundary operators can all be seen as acting in combination upon the solution for the ocean circulation, and producing the inhomogeneities or inputs which are, again, the subsurface forcing, initial values, surface forcing and any values at open boundaries. The combined operator is nonsingular if there exists a unique and analytically satisfactory – say, continuously differentiable – solution for each set of analytically satisfactory inputs. If the operator is nonsingular, then there is a welldeﬁned and unique inverse operator. From the mathematical perspective, the solution is the action of the inverse operator on the inputs. Computing this action is, in the inﬁnite wisdom of convention, “forward ocean modeling”. Characterizing our knowledge of ocean circulation as solutions of wellposed, mixed initialvalue boundaryvalue problems does not correspond to our real experience of the ocean. Ship surveys, moored instruments, buoys drifting freely on the ocean surface or ﬂoating freely below the surface, and earth satellites orbiting above cannot observe continuous ﬁelds throughout an ocean region, even for one instant. Yet these data, after control for quality, are in general far more reliable than either the parameterizations of turbulence in the dynamics or the crudely interpolated forcing ﬁelds, initial values and boundary values. The qualitycontrolled data belong to any rational concept of a model, and the set of mathematical elements that deﬁnes a model is readily extended to include them. Speciﬁcally, functionals corresponding to methods of measurement, and numbers corresponding to measurements of quantities in the real ocean (such as velocity components, temperature, density and the like), may be added to the set. Each functional maps a circulation ﬁeld into a single number. For example, monthlymean sea level at a coastal station deﬁnes a kernel or integrand which selects sea level from the many circulation variables, which has a rectangular time window of one month and which is sharply peaked at the coastal station. Note that the new mathematical elements
Preface
xv
include both additional operators (the measurement functionals), and additional input (the data). It is always assumed, but almost never proved, that the operator for the original model is nonsingular. It can always be assumed that applying the measurement functionals to the unique solution of the original forced, initialboundary value problem does not produce numbers equal to the real data. The extended operator can therefore have no inverse, and so must be singular. It seems natural, even a compulsion (Reid, 1968), to determine the ocean circulation as some uniquelydeﬁned bestﬁt to the extended inputs (forcing, initial, boundary and observed). The singular extended operator then has a generalized inverse operator, and the bestﬁt ocean circulation is the action of the generalized inverse on the extended inputs. This book outlines the theoretical and practical computation of the action, for bestﬁts in the sense of weighted leastsquares. The practical computations will be only numerical approximations, so the theme of the book should therefore be expressed as “inverting numerical models and observations of the ocean and atmosphere in a generalized sense”. Abandoning precision for brevity, the theme is “inverse modeling the ocean and atmosphere”. How, then, does inverse modeling meet the needs of oceanographers and meteorologists? The bestﬁt circulation is clearly an analysis, an optimal dynamical interpolation in fact, of the observations. All the ﬁelds coupled by the dynamics are analyzed, even if only some of them are observed. The leastsquares ﬁt to all the information, observational and dynamical, yields residuals in the equations of motion as well as in the data, and these residuals may be interpreted as inferred corrections to the dynamics or to the inputs. There are emerging techniques that can in principle distinguish between additive errors in dynamics and internal forcing, but these techniques are so new and unproven that it would be premature, even by the standards of this infant discipline, to include them here. Empirical parameters may also be tuned to improve the analysis. (The tuning game, sometimes described as a “ﬁddler’s paradise” [Ljung and S¨oderstr¨om, 1987], is outlined here.) The conditioning or sensitivity of the ﬁt to the inputs, as revealed during the construction of the generalized inverse, quantiﬁes the effectiveness of the observing system. The natural choices for the weights in the best ﬁt are inverses of the covariances of the errors in all the operators and inputs. These covariances must be stipulated by the inverse modeler. They accordingly constitute, along with stipulated means, a formal hypothesis about the errors in the model and observations. The minimized value of the ﬁtting criterion or penalty functional yields a signiﬁcance test of that hypothesis. For linear leastsquares, the minimal value is the χ 2 variable with as many degrees of freedom as there are data, provided the hypothesized means and covariances are correct. A failed signiﬁcance test does discredit the analyzed circulation and also any concomitant assessment of the observing system, but does not end the investigation: detailed examination of the residuals in the equations, initial conditions, boundary conditions and data can identify defects in the model or in the observing system. Thus model development can proceed in an orderly and objective fashion. This is not to deny the crucial roles of astute and inspired insight in oceanic and
xvi
Preface
atmospheric model development as in all of science; it is rather to advocate a minimal level of organization especially when inspiration is failing us, as seems to be the case at present. In spite of an explicit emphasis here on time dependence, the spirit of this approach is close to geophysical inverse theory (see for example Parker, 1994), speciﬁcally the estimation of permanent strata in the solid earth using seismic data. The retrieval of instantaneous vertical proﬁles of atmospheric temperature and moisture using multichannel microwave soundings from satellites (see for example Rodgers, 2000) bears a striking resemblance to the seismic problem, and is indeed both named and practised as inverse theory. Yet the context here – timedependent oceanic and atmospheric circulation – is so different that to call it inverse theory seems almost misleading. Inverse modeling is but one formulation of the vaguely deﬁned activity known as “data assimilation”. The most widely practised form of oceanic or atmospheric data assimilation involves interpolating ﬁelds at one time, for subsequent use as initial data in a model integration which may even be a genuine forecast. Once nature has caught up with the forecast, the latter serves as a ﬁrstguess or “background” ﬁeld for the next synoptic analysis. As might be imagined, this cycle of synoptic analysis and forecasting is a major enterprise at operational centers, and is very extensively developed for meteorological applications. Characterization of operational systems for observing the weather, in particular studying the statistics of observational errors, has been and remains the subject of vast investigation. Comprehensive references may be found at appropriate places in the following chapters, but that description of operational detail will not be repeated here. Nor will the intricate, “diagnosticallyconstrained” multivariate forms of synoptic interpolation be discussed in detail. Geostrophy, for example, is an approximate diagnostic constraint on synoptic ﬁelds of velocity and pressure. The emphasis instead will be on elaborating the new data assimilation schemes that could be consistently described as nonsynoptic, “prognosticallyconstrained” interpolation. The unapproximated law of conservation of momentum, for example, is a prognostic constraint. Again, the nature of this latter activity is so different in technique and broader in scope, in comparison with the conventional cycle of synoptic analysis and forecasting, that to call these new schemes “data assimilation” seems to be misleading yet again. The name “inverse modeling” is chosen, for better or worse. What else has been left out here? Monte Carlo methods are immensely appealing in any application, and data assimilation is no exception. Sample estimates of means and covariances of circulation ﬁelds may be generated from repeated forward integrations of a model driven by suitably constructed pseudorandom inputs. The sample moments of the solutions are then used for conventional synoptic interpolation. The calculus of variations is not required. These assimilation methods, especially “ensemble Kalman ﬁltering”, are highly competitive with variational inverse methods in terms of development effort and computational efﬁciency, but are even more immature and so are mentioned only brieﬂy. The very basics of statistical simulation and Monte Carlo methods in general are outlined in these chapters.
Preface
xvii
The reader should not be discouraged by the technical deﬁnition of inverse modeling given in previous paragraphs. The calculus of several variables, a rudimentary knowledge of partial differential equations and the same numerical analysis used to solve the forward model are enough mathematics for the computation of generalized inverses. Abstraction is restricted to the one place in this book where an elegant expression of generality is of real beneﬁt. The Hilbert Space analysis sketched in Chapter 2 exposes the geometrical structure of the generalized inverse, and explains the efﬁciency of the concrete algorithms developed in Chapter 1. The geometrical interpretation is a straightforward adaptation of the theory of Laplacian spline interpolation. A beautiful treatment of Lsplines may be found in an applied meteorology journal (Wahba and Wendelberger, 1980), to the eternal credit of the authors, reviewers and editors. Any temptation to make use of the Hilbert Space machinery for abstract deﬁnitions of adjoint operators is easily resisted, as such abstraction offers no real insight into the problem of interest. The adjoint operators arise naturally when the elementary calculus of variations is used to derive the classic Euler–Lagrange conditions for the weighted, leastsquares bestﬁt. Unlike the Hilbert Space deﬁnition of an adjoint operator, the variational calculus need not be preceded by a linearization of the dynamics and measurement functionals. This ﬂexibility leads to critically important alternatives for iterative solution techniques that are linear. It is essential to distinguish the formulative and interpretive aspects of inverse modeling from its mathematical aspects. Leastsquares may be used to estimate any quantity, but it is the estimator of maximum likelihood for Gaussian or normal random variables. Such variability can reasonably be expected in the ocean and atmosphere, on the synoptic scale and larger, away from transient and semipermanent fronts, and in variables not subject to phase changes. Leastsquares is especially attractive from a mathematical perspective, since it leads to linear conditions for the best ﬁt when the constraints are also linear. The linearity of the extremal conditions permits powerful analyses which yield efﬁcient solution methods. There are many leastsquares algorithms, such as optimal interpolation, Kalman ﬁltering, ﬁxedinterval smoothing, and representers. The relationships between these statistical, controltheoretic and geometrical approaches are explained in this book. Aside from unifying the mathematics, recognizing the mathematical relationships facilitates the identiﬁcation of scientiﬁc assumptions. For example, if the data were collected in much less time than the natural scales of evolution of the dynamics and the internal forcing, then there would be little to gain by assuming that there are errors in the dynamics or internal forcing. It would sufﬁce to admit errors only in the initial conditions, surface forcings, open boundary values and data. This assumption massively reduces the ﬁnite dimension of the “state space” for the numerical model, by eliminating those variable ﬁelds or “controls” deﬁned both throughout the ocean region and throughout the time interval of interest. Boundary values, initial values and empirical parameters would be retained as controls. The reduced state or “control” space may be sufﬁciently small that a conventional
xviii
Preface
gradient search for a minimum in the state space is feasible. The condition of the ﬁt in state space determines the efﬁciency of the search. It is, however, becoming increasingly necessary to consider time intervals during which the dynamical errors are bound to become signiﬁcant. That is, the initial conditions would be ineffective as controls for guiding the model solution towards the later data. Using distributed controls, that is, admitting errors in the dynamics throughout space and time, leads to huge numbers of computational degrees of freedom. (There are in general far fewer statistically independent degrees of freedom, but these are not readily identiﬁed. Indeed, the methods developed here serve to identify them.) Hence there could be no prospect of a wellconditioned search in the control space or in the equivalent state space. The power of the methods described in these chapters is that they identify a huge subspace of controls (known as the null space) having exactly no inﬂuence on guiding the solution towards the data. The methods restrict the search for optimal controls to those lying entirely in the comparatively tiny, orthogonal complement of the null space (known as the data subspace). Again, as in the choice of a leastsquares estimator, there is an interplay between scientiﬁc formulation and mathematical technique. The two should nonetheless always be carefully distinguished. As a ﬁnal example of the distinction and interplay between scientiﬁc formulation and mathematical manipulation, consider errors in models of smallscale ﬂows. As already implied, these errors are likely to be highly intermittent or nonGaussian. Thus, inversions of observations collected in mixing fronts and jets, or in free convection, or during phase changes, will require estimators other than least squares. Only bruteforce minimization techniques, such as simulated annealing or Monte Carlo methods in general, appear to be available for most estimators. On the other hand, multiprocessor computer architecture may favor bruteforce inversion. These bruteforce techniques will be mentioned here, but only brieﬂy, since by their nature regrettably little is known about them. The content of this book closely follows an upperlevel graduate course for physical oceanography students at Oregon State University. Their preparation includes r graduate courses in ﬂuid dynamics, geophysical ﬂuid dynamics and ocean circulation theory; r a graduate course in numerical modeling of ocean circulation; r a graduate course in time series analysis including Gauss–Markov smoothing or “objective analysis”; r graduate courses in ordinary and partial differential equations, computational linear algebra and numerical methods in general; r FORTRAN and basic UNIX skills; r or an equivalent preparation in atmospheric science. The curriculum does not require great depth or fresh familiarity with all of the above material. The following would sufﬁce.
Preface
xix
1. Some minimal exposure to hydrodynamics, preferably in a rotating reference frame, including approximations such as hydrostatic balance, the shallowwater equations and geostrophic balance. The wellknown texts by Batchelor (1973), Pedlosky (1987), Gill (1982), Holton (1992) and Kundu (1990) may be consulted. Graduate students in physics or mechanical or civil engineering would have no problem with the curriculum, although some jargon may cause them to glance at a text in oceanography or meteorology. 2. The knowledge that oceanic and atmospheric circulation models are expressed as partial differential equations (pdes) that may be numerically integrated, most simply using ﬁnite differences. The text by Haltiner and Williams (1980) on numerical weather prediction is very useful. 3. Access to Stakgold’s classic (1979) text on boundary value problems. The theoretical notions most useful here are (i) odes and pdes can only have wellbehaved solutions if precisely the right number of initial and boundary value conditions are provided and (ii) the solution of such wellposed problems for linear odes and pdes can be expressed using a Green’s function or inﬂuence function. As for computational linear algebra and numerical methods in general, the synopses in Press et al. (1986) are very useful. 4. Comfort with the very basics of probability and statistics, including random variables, means, covariances and minimumvariance estimation. Again, the synopses in Press et al. (1986) make a good ﬁrst reading. 5. As much FORTRAN as can be learned in a weekend. The content of the Preamble, and of each of the six chapters and the two appendices, is outlined on their ﬁrst pages. The Preamble attempts to communicate the nature of variational ocean data assimilation, or any other assimilation methodology, through a commonplace application of basic scientiﬁc method to marine biology. The example might seem out of context, and indeed it is, but that underscores the universality and long history of the approach advocated here. Its arrival in the context of oceanic and atmospheric circulation has of course been delayed by the fantastic mathematical and computational complexity of circulation models. The Preamble includes a “data assimilation checklist”, which the student or researcher is encouraged to consult regularly. Chapter 1 is the irreducible introduction to variational assimilation with dynamical models; a “toy” model consisting of a single linear wave equation with one space dimension serves as an illustration. Chapter 2 complements the controltheoretic development of Chapter 1 with geometrical and statistical interpretations; analytical considerations essential to the physical realism of the inverse solutions are introduced. Chapter 3 addresses efﬁcient construction of the inverse and its error statistics, and introduces iterative techniques for coping with nonlinearity. Chapter 4 surveys alternative algorithms for linear leastsquares assimilation, and for assimilation with nonlinear or nonsmooth models or with nonlinear measurement functionals. Difﬁculties to be expected with nonlinear techniques are outlined – proven remedies are still
xx
Preface
lacking. Chapter 5 reviews largescale geophysical ﬂuid dynamics, discusses several real oceanic and atmospheric inverse models in detail, and concludes with notes on a selection of contemporary efforts, both research and operational. Chapter 6 applies inverse methods to forward models based on singular operators. The material in this book can be presented in thirty onehour lectures. An overhead projector is a great help: minimaltext, mathonly, largefont overhead transparencies allow the audience to listen, rather than transcribe incorrectly. The overheads are available as TEX source ﬁles via an anonymous ftp site (ftp.oce.orst.edu, dist/bennett/class/overheads). Students should be able to begin the computing exercises in Appendix A after studying the ﬁrst four sections of Chapter 1. The inverse tidal model of §5.2 in Chapter 5 is accessible after studying Chapters 1 and 2. The nonlinear inverse models of tropical cyclones and ENSO in §5.3–5.5, and the accelerated algorithms used in their construction, require a study of Chapter 3. The complete variational equations for the tropical cyclone inversion may be found in Appendix B. A ﬁrst reading of Chapter 4 is assumed in §5.6, the survey of contemporary applications of advanced assimilation with oceanic and atmospheric data. The research monograph by Bennett (1992) contains almost all of the theoretical development found here, but none of the guidelines for implementation and few case studies with real data or real arrays. Certain advanced theoretical considerations, such as Kalman ﬁlter pathology in the equilibrium limit and continuous families of representers for excess boundary data, are only brieﬂy mentioned here if at all, but may be found in the earlier monograph. There has been a rapid growth in the literature of nonsynoptic data assimilation during the last decade. A full literature survey would be impractical and of doubtful value as so much work has been highly applicationspeciﬁc. Shorter but very useful survey articles include Courtier et al. (1993); Anderson, Sheinbaum and Haines (1996) and Fukumori (2001); for collections of expository papers and applications see MalanotteRizzoli (1996), Ghil et al. (1997) and Kasibhatla et al. (2000). The lastmentioned is noteworthy for its interdisciplinary range, and also for a set of exercises on various assimilation techniques. The major text by Wunsch (1996) principally develops in great detail the timeindependent inverse theory for steady ocean circulation, using a ﬁnitedimensional formulation which certainly complements the analytical development here and which may be the more accessible for being ﬁnitedimensional. On the other hand the essential mathematical condition of the inverse problem is established at the analytical or continuum level, and the “look and feel” of geophysical ﬂuid dynamics is retained by an analytical formulation. Inverse modeling suffers not so much from the lack of good data, credible models and adequate computing resources, as from a lack of experience. This book is intended to be of assistance to the generation of investigators who, it is hoped, will acquire that experience. Monterey, June 2001
Acknowledgements
Many of my collaborators and colleagues over the last two decades have inﬂuenced this book, most recently Boon Chua, Gary Egbert and Robert Miller. The exercises in Appendix A were devised and constructed by Boon Chua and HansEmmanuel Ngodock. The book is based upon lecture notes for two summer schools on inverse methods and data assimilation, held at Oregon State University in 1997 and 1999. Numerous suggestions from the summer school participants led to signiﬁcant improvement of the notes. Successive versions were all patiently typed by Florence Beyer, and the book manuscript was created from the notes by William McMechan. Permission to reproduce copyrighted ﬁgures was received from: InterResearch (Fig. P.1.1), Elsevier (Fig. 5.2.7), the American Geophysical Union (Fig. 5.2.8), Nature (Fig. 5.3.5), SpringerVerlag (Figs. 5.4.1, 5.4.2, 5.4.3, 5.4.5), and the American Meteorological Society (Figs. 5.5.1, 5.5.2, 5.5.3, 5.5.4, 5.5.5, 5.5.6, 5.5.7, 5.5.8). Dudley Chelton kindly provided Fig. 5.2.5. All the other ﬁgures were drafted by David Reinert. Thanks for the cover illustrations are owed to Michael McPhaden and Linda Stratton of the NOAA Paciﬁc Marine Environmental Laboratory. The support of the National Science Foundation (OCE9520956) is also acknowledged. It has been a pleasure to work with Cambridge University Press: my editors Alan Harvey and Matt Lloyd; also Nicola Stern, Susan Francis, Jo Clegg, and Frances Nex. The manuscript was completed at the Ofﬁce of Naval Research Science Unit in the Fleet Numerical Meteorology and Oceanography Center, under the auspices of the Visiting Scientist Program of the University Corporation for Atmospheric Research. I am grateful to Oregon State University for an extended leave of absence, to Manuel
xxi
xxii
Acknowledgements
Fiadeiro at ONR and Michael Clancy at FNMOC, to Meg Austin and her team at the UCAR VSP Ofﬁce for their friendly efﬁciency, and to Captain Joseph Swaykos, USN and his entire staff, for their hospitality during an educational and productive stay at “Fleet Numerical”.
Internet sites mentioned in this book: bioloc.coas.oregonstate.edu/pictures/gallery2/index.html ftp.oce.orst.edu, cd/dist/chua/IOM/IOSU, 216 ftp.oce.orst.edu, cd/dist/bennett/class, 146, 195 ftp.oce.orst.edu, cd/dist/bennett/class/overheads, xx www.gfdl.gov/~ smg/MOM/MOM.html, 167 www.polar.gsfc.nasa.gov 166 www.fnmoc.navy.mil, 166 www.pac.dfompo.gc.ca/sci/osap/projects/plankton/zoolab e.htm#Copepod www.oce.orst.edu/po/research/tide/global.html, 136 www.pmel.noaa.gov/togatao/home.html, 156 http://diadem.nersc.no/project, 169 www.units.it/~ mabiolab/set previous.htm, 1 Note: oce.orst.edu changes to coas.oregonstate.edu in 2002. The publisher has used its best endeavors to ensure that the URLs for external websites referred to in this book are correct and active at the time of going to press. However, the publisher has no responsibility for the websites and can make no guarantee that a site will remain live or that the content is or will remain appropriate.
Preamble
An ocean data assimilation system in miniature The pages of this book are ﬁlled with the mathematics of oceanic and atmospheric circulation models, observing systems and variational calculus. It would only be natural to ask: What is going on here, and is it really new? The answers are “regression” and hence “no”: almost every issue of any marine biology journal contains a variational ocean data assimilation system in miniature.
P.1
Linear regression in marine biology
The article “Repression of fecundity in the neritic copepod Acartia clausi exposed to the toxic dinoﬂagellate Alexandrium lusitanicum: relationship between feeding and egg production”, by J¨org Dutz, appeared in Marine Ecology Progress Series in 1998. Dinoﬂagellates are a species of phytoplankton, or small plantlike creatures. The genus Alexandrium (www.units.it/~ mabiolab/set previous.htm, click on ‘Toxic microalgae’) produces toxins which rise through the food web to produce paralytic shellﬁsh poisoning in a variety of hydrographical regions, ranging from temperate to tropical. Zooplankton, or small animallike creatures (www.ios.bc.ca/ios/plankton/ios tour/zoop lab/copepod.htm), graze on these dinoﬂagellates. The effect of the toxins on the grazers naturally arises. Dutz (1998) fed toxinbearing Alexandrium lusitanicum and toxinfree Rhodomonas baltica (bioloc.coas.oregonstate.edu/baltica.jpg) to females of the copepod Acartia clausi in controlled amounts, and measured the fecundity or gross growth
1
2
Preamble
Figure P.1.1 Gross growth efﬁciency of Acartia clausi versus food supply. Solid circles: nontoxic Rhodomonas baltica; open circles: toxic Alexandrium lusitanicum (after Dutz, 1998).
0.5 0.4 0.3 0.2 0.1 0
0
400
800
1200
1600
efﬁciency in terms of total carbon production. He found that the grazers were not killed, and they continued to lay eggs. However, their fecundity was affected: see Fig. P.1.1. Note the controlled food concentration (abscissa x) with ﬁve values: 200, 400, 800, 1200, and 1600 µg C l−1 . Fecundity is not inﬂuenced by the supply of nontoxic Rhodomonas (solid circles), but is clearly reduced as the supply of toxic Alexandrium (open circles) increases. The gross growth efﬁciencies (ordinate y) in the latter case are respectively: 0.23, 0.21, 0.18, 0.14, 0.10 (Dutz, 1998; Table 2). The error bars indicate Dutz’ maximum and minimum estimates. A straight line clearly ﬁts the Alexandrium data well. The regression parameters are: a = 0.25, b = 9.2 × 10−5 , r 2 = 0.997, F1, 3 = 355, P < 0.0005. A brief review of linear regression is in order. The data are M ordered pairs: (xm , ym ), 1 ≤ m ≤ M. The model is ym = α + βxm + m ,
(P.1.1)
where α and β are unknown constants, while m is a random variable with mean and covariance 2 σ , n=m 2 (P.1.2) Em = 0, E(m n ) = σ δnm = 0, n = m. The error m is an admission of measurement error, and of the unrepresentativeness of a linear relationship. Note that the model consists of an explicit functional form (here, a linear relationship), together with probabilistic statements (here, mean and covariance) about the error in the form. We seek an estimate (here, a regression line): yˆ = a + bx,
(P.1.3)
where a and b are to be chosen. As an estimator, let us choose a uniformly weighted sum of squared errors: W SS E = σ −2
M m=1
(ym − a − bxm )2 .
(P.1.4)
P.1 Linear regression in marine biology
3
A value for σ may be inferred from the error bars in Fig. P.1.1. It is easily shown that W SS E is minimal if a and b satisfy the normal equations: 1 x a y = , (P.1.5) b xy x x2 M xm . Note where the overbar denotes the arithmetic mean, for example x = M −1 m=1 −2 that (P.1.5) is independent of the uniform weight σ . These equations are of course trivially solved for a and b. The following statements may be made about the ﬁrst and second moments of the solution: Ea = α, E(a − α)2 =
x 2σ 2 M(x 2 − (x)2 )
Eb = β,
,
E(b − β)2 =
σ2 M(x 2 − (x)2 )
.
(P.1.6)
Moreover, a, b and yˆ m are normally distributed around α, β and ym respectively. Note that the error variances in (P.1.6) are O(M −1 ). In addition to the posterior error estimates (P.1.6), there are signiﬁcance test statistics such as the varianceratio or F test: M
F1,M−2 =
(ym − y)2
m=1 M
,
(P.1.7)
(ym − yˆ m )
m=1
where yˆ m ≡ axm + b. The numerator is the total variance of the data; the denominator is the total variance of the residuals for the regression line (P.1.3). Note that (P.1.7) is independent of σ 2 . The subscripts 1 and M − 2 indicate the number of degrees of freedom in the denominator and the numerator, respectively. The value of F here is 355; accordingly the probability P of the null hypothesis (α = β = 0) being true is less than 0.05%. In other words it is highly credible that grazing on Alexandrium lusitanicum does repress the fecundity of Acartia clausi.
Exercise P.1.1 An alternative test statistic is provided by the weighted denominator in (P.1.7): resW SS E = σ −2
M
(ym − yˆ m )2
m=1
∼
2 χM ,
as
M → ∞.
(P.1.8)
2 2 = M, varχ M = 2M. Calculate (P.1.8) using Dutz’ data, and draw Verify that Eχ M conclusions.
If the data had suggested it, Dutz could have considered quadratic regression: ym = α + βxm + γ xm2 + m , Em = 0,
E(n m ) = σ 2 δnm .
(P.1.9)
4
Preamble
Figure P.1.2 On the left: the parabola of leastsquares best ﬁt to four data points, which are shown as solid circles. The abscissa values for the data (see the tick marks on the abscissa in the zoom on the right) are illchosen. As a result, the leastsquares best ﬁt is clearly illconditioned. The abscissa itself would be a more sensible ﬁt to the data.
The estimate would be yˆ = a + bx + cx 2 . The estimator would again be (P.1.4), for which the normal equations are y 1 x x2 a x x 2 x 3 b = x y . c x2y x2 x3 x4
(P.1.10)
(P.1.11)
Suppose for simplicity that x = x 3 = 0 (these are at our disposal). Then the system (P.1.11) is illconditioned; that is, the solution (a, b, c) is highly sensitive to the inhomogenity on the righthand side if x 4 /(x 2 )2 1. This ratio is also at our disposal. Just such a situation is sketched in Fig. P.1.2. The best ﬁt to the four data points is a deep parabola, yet the most sensible ﬁt would be the abscissa itself (y = 0). In conclusion, the stability of the estimate (P.1.10) is controlled by the choice of abscissa values xm , 1 ≤ m ≤ M. P.2
Data assimilation checklist
The preceeding elementary application of linear regression in marine biology has every aspect of an “ocean data assimilation system”: see the following checklist.
Data assimilation checklist INP UTS (i) There is an observing system, consisting of measurements of gross growth efﬁciency at selected food concentration levels.
P.2 Data assimilation checklist
5
(ii) There are dynamics, expressed here as (P.1.1), the explicit general solution of the differential equation d2 y = 0, dx2
(P.2.1)
plus measurement errors m , 1 ≤ m ≤ M. The values α, β indicated in (P.1.1) for the regression constants a, b are the “true” values. (iii) There is an hypothesis (P.1.2) about the distribution of errors m around the true regression line. (iv) There is an estimator, here the uniformly weighted sum of squared errors (P.1.4). (v) There is an optimization algorithm, here the normal equations (P.1.5) which would, in the general case of N th order polynomial regression, be robustly solved using the singular value decomposition.
OUTPUTS (vi) There is an estimate of the state, here the regression line (P.1.3) with values of a and b obtained from the normal equations (P.1.5). (vii) There are estimates of data residuals and dynamical residuals. Here the two types of residual are indistinguishable; both are in fact given by ym − yˆ m . (viii) There are posterior error statistics, here the means and variances (P.1.6) for a − α and b − β. (ix) There is an assessment of the array or observing system. Here it is the conditioning of the normal matrix, and is determined by the choices of food concentrations xm , 1 ≤ m ≤ M. (x) There are test statistics, here the Fvariable (P.1.7) and χ 2 variable (P.1.8). These indicate the credibility of the hypothetical model, and thus the credibility of the derived posterior error statistics. (xi) There are indications for model improvement. Here, however, the indication is that the linear model is so credible that a quadratic model (P.1.10) is unnecessary. Variational assimilation of El Ni˜no data from the tropical Paciﬁc, into a coupled intermediate model of the ocean and atmosphere, is described in §5.5. The checklist reads as follows.
INPUTS (i) The observations are monthlymean and ﬁveday mean values of Sea Surface Temperature (SST, or T (1) ), the depth of the 20◦ isotherm (Z 20) and surface winds (u a , v a ), at the TOGA–TAO moorings, from April 1994 to May 1998. (ii) The dynamics are those of an intermediate coupled model after Zebiak and Cane (1987); the thermodynamics of the upper oceanic layer and the coupling through the wind stress are nonlinear. Otherwise the oceanic and atmospheric dynamics are those of linearized shallowwater waves.
6
Preamble
(iii) The hypothesis consists of means and autocovariances of errors in the dynamics, in the initial conditions and in the data. (iv) The estimator is the combined, spaceintegrated and timeintegrated weighted squared error. (v) The optimization algorithm is the iterated, indirect representer algorithm for solving the nonlinear Euler–Lagrange equations.
OUTPUTS (vi) There are estimates of spacetime ﬁelds of surface temperature, currents, thermocline depths and surface winds. (vii) There are corresponding space–time ﬁelds of minimal residuals in the dynamics, initial conditions and data. (viii) There are space–time covariances of errors in the optimal estimates of the coupled circulation. (ix) These are assessments of the efﬁciency of the monthlymean TOGA–TAO system for observing the “weak” dynamics of the coupled model, that is, observing the intermediate dynamics subject to the hypothesized error statistics. (x) The reduced estimator is a χ 2 variable for testing the hypothesized error moments (they were found to lack credibility). (xi) The dominance of the minimal residual in the upperocean thermodynamic balance indicates that it would serve no purpose to hypothesize increased variances for the dynamical errors: the lowresolution intermediate dynamics should be abandoned in favor of a fullystratiﬁed, highresolution, Primitive Equation model. Variational data assimilation, or generalized inversion of dynamical models and observations, is really no more than regression analysis. The novelty lies in the mathematical and physical subtlety of realistic dynamics, in the complexity of the hypotheses about the multivariate random error ﬁelds, and in the sheer size of modern data sets. The novelty also lies in the emergence of powerful and efﬁcient optimization algorithms, which allow us to test our models in the same way that all other scientists test theirs.
Chapter 1 Variational assimilation
Chapter 1 is a minimal course on assimilating data into models using the calculus of variations. The theory is introduced with a “toy” model in the form of a single linear partial differential equation of ﬁrst order. The independent variables are a spatial coordinate, and time. The wellposedness of the mixed initialboundary value problem or “forward model” is established, and the solution is expressed explicitly with the Green’s function. The introduction of additional data renders the problem illposed. This difﬁculty is resolved by seeking a weighted leastsquares best ﬁt to all the information. The ﬁtting criterion is a penalty functional that is quadratic in all the misﬁts to the various pieces of information, integrated over space and time as appropriate. The bestﬁt or “generalized inverse” is expressed explicitly with the representers for the penalty functional, and with the Green’s function for the forward model. The behavior of the generalized inverse is examined for various limiting choices of weights. The smoothness of the inverse is seen to depend upon the nature of the weights, which will be subsequently identiﬁed as kernel inverses of error covariances. After reading Chapter 1, it is possible to carry out the ﬁrst four computing exercises in Appendix A.
1.1
Forward models
1.1.1
Wellposed problems
Mechanics is captured mathematically by “wellposed problems”. The mechanical laws for particles, rigid bodies and ﬁelds are with few exceptions expressed as ordinary or partial differential equations; data about the state of the mechanical system are provided
7
8
1. Variational assimilation
in initial conditions or boundary conditions or both. The collection of general equations and ancillary conditions constitute a “wellposed problem” if, according to Hadamard (1952; Book I) or Courant and Hilbert (1962; Ch. III, §6): (i) a solution exists, which (ii) is uniquely determined by the inputs (forcing, initial conditions, boundary conditions), and which (iii) depends continuously upon the inputs. Classical particles and bodies move smoothly, while classical ﬁelds vary smoothly so only differentiable functions qualify as solutions. The repeatability of classical mechanics argues for determinism. The classical perception of only ﬁnite changes in a ﬁnite time argues for continuous dependence. Illposed problems fail to satisfy at least one of conditions (i)–(iii). They cannot be solved satisfactorily but can be resolved by generalized inversion, which is the subject of this chapter. Inevitably, wellposed problems are also known as “forward models”: given the dynamics (the mechanical laws) and the inputs (any initial values, boundary values or sources), ﬁnd the state of the system. In this ﬁrst chapter, an example of a forward model is given; the uniqueness of solutions is proved, and an explicit solution is constructed using the Green’s function. That is, the wellposedness of the forward model is established. 1.1.2
A “toy” example
The following “toy” example involves an unknown “ocean circulation” u = u(x, t), where x, t and u are real variables. The “ocean basin” is the interval 0 ≤ x ≤ L, while the time of interest is 0 ≤ t ≤ T : see Fig. 1.1.1. The “ocean dynamics” are expressed as a linear, ﬁrstorder partial differential equation: ∂u ∂u +c =F (1.1.1) ∂t ∂x for 0 ≤ x ≤ L and 0 ≤ t ≤ T , where c is a known, constant, positive phase speed. The inhomogeneity F = F(x, t) is a speciﬁed forcing ﬁeld; later it will become known as the prior estimate of the forcing. An initial condition is u(x, 0) = I (x)
(1.1.2)
for 0 ≤ x ≤ L, where I is speciﬁed. A boundary condition is u(0, t) = B(t) for 0 ≤ t ≤ T , where B is speciﬁed.
(1.1.3)
1.1 Forward models
9
t
Figure 1.1.1 Toy ocean basin.
T
x
L
0
1.1.3
Uniqueness of solutions
To determine the uniqueness of solutions (Courant and Hilbert, 1962) for (1.1.1), (1.1.2) and (1.1.3), let u 1 and u 2 be two solutions for the same choices of F, I and B. Deﬁne the difference v ≡ u1 − u2.
(1.1.4)
∂v ∂v +c =0 ∂t ∂x
(1.1.5)
v(x, 0) = 0
(1.1.6)
v(0, t) = 0
(1.1.7)
Then
for 0 ≤ x ≤ L and 0 ≤ t ≤ T ;
for 0 ≤ x ≤ L, and
for 0 ≤ t ≤ T . Multiplying (1.1.5) by v and integrating over all x yields d 1 dt 2
L v 2 d x = −c
1 2 v 2
0
x=L x=0
c = − v(L , t)2 , 2
(1.1.8)
using the boundary condition (1.1.7). Integrating (1.1.8) over time from 0 to t yields 1 2
L
1 v (x, t) d x = 2
L
2
0
c v (x, 0) d x − 2
t v 2 (L , s) ds.
2
0
(1.1.9)
0
The righthand side (rhs) of (1.1.9) is nonpositive, as a consequence of the initial condition (1.1.6). Hence v(x, t) = 0,
(1.1.10)
10
1. Variational assimilation
that is, u 1 (x, t) = u 2 (x, t)
(1.1.11)
for 0 ≤ x ≤ L and 0 ≤ t ≤ T . So we have established that (1.1.1), (1.1.2) and (1.1.3) have a unique solution for each choice of F, I and B. 1.1.4
Explicit solutions: Green’s functions
We may construct the solution explicitly, using the Green’s function (Courant and Hilbert, 1953) or fundamental solution γ for (1.1.1)–(1.1.3). Let γ = γ (x, t, ξ, τ ) satisfy ∂γ ∂γ −c = δ(x − ξ )δ(t − τ ), ∂t ∂x where the δs are Dirac delta functions, and 0 ≤ ξ ≤ L , 0 ≤ τ ≤ T . Also, −
(1.1.12)
γ (L , t, ξ, τ ) = 0
(1.1.13)
γ (x, T, ξ, τ ) = 0
(1.1.14)
for 0 ≤ t ≤ T , and for 0 ≤ x ≤ L.
Exercise 1.1.1 (a) Verify that γ (x, t, ξ, τ ) = δ(x − ξ − c(t − τ ))H (τ − t)
(1.1.15)
for 0 ≤ x < L, 0 ≤ t ≤ T , where H is the Heaviside unit step function. (b) Show that T u(ξ, τ ) = u F (ξ, τ ) ≡
L d x γ (x, t, ξ, τ )F(x, t)
dt 0
0
L +
T d x γ (x, 0, ξ, τ )I (x) + c
0
dt γ (0, t, ξ, τ )B(t).
(1.1.16)
0
Relabeling (1.1.16) yields T u F (x, t) =
L dξ γ (ξ, τ, x, t)F(ξ, τ )
dτ 0
0
L T + dξ γ (ξ, 0, x, t)I (ξ ) + c dτ γ (0, τ, x, t)B(τ ), 0
0
(1.1.17)
1.1 Forward models
11
which is an explicit solution for the “forward model”. It is also the prior estimate or “ﬁrstguess” or “background” for u. Note 1. By inspection, u F depends continuously upon changes to F, I and B; if these change by O(), so does u F . Note 2. We actually require I (0) = B(0), or else u F is discontinuous across the phase line x = ct, for all t. We conclude that the forward model (1.1.1)–(1.1.3) is wellposed. Any additional information would overdetermine the system, and a smooth solution would not exist.
Exercise 1.1.2 Code the ﬁnitedifference equation
u k+1 = u kn − c(t/x) u kn − u kn−1 + t Fnk , n
(1.1.18)
where u kn = u(nx, kt), etc. Perform some numerical integrations. Derive and verify experimentally the Courant–Friedrichs–Lewy stability criterion (Haltiner and Williams, 1980).
Exercise 1.1.3 Slow, onedimensional viscous ﬂow u = u(x, t) is approximately governed by ∂u ∂p ∂ 2u = ν 2 − ρ −1 , (1.1.19) ∂t ∂x ∂x where ν is the uniform kinematic viscosity, ρ is the uniform density and p = p(x, t) is the externally imposed pressure gradient. Consider an inﬁnite domain: −∞ < x < ∞, and a ﬁnite time interval: 0 < t < T . A suitable initial condition is u(x, 0) = I (x).
(1.1.20)
Assume that both ∂ p/∂ x and I vanish as x → ∞. (a) Derive the following energy integral when both ∂ p/∂ x and I vanish everywhere: d 1 dt 2
∞
∞ u d x = −ν 2
−∞
−∞
∂u ∂x
2 d x.
(1.1.21)
Hence prove that there is at most one solution for each choice of p and I . (b) Show that the solution of (1.1.19), (1.1.20) is u(x, t) = −ρ −1
T 0
∞ + −∞
∞ dξ φ(ξ, τ, x, t)
dτ −∞
dξ φ(ξ, 0, x, t)I (ξ ),
∂p (ξ, τ ) ∂x (1.1.22)
12
1. Variational assimilation
where the Green’s function or fundamental solution φ(x, t, ξ, τ ) satisﬁes −
∂φ ∂ 2φ = ν 2 + δ(x − ξ )δ(t − τ ), ∂t ∂x
(1.1.23)
subject to φ(x, T, ξ, τ ) = 0
(1.1.24)
φ(x, t, ξ, τ ) → 0
(1.1.25)
for −∞ < x < ∞, and
as x → ∞. (c) Verify that −(x−ξ )2
H (τ − t)e 2ν(τ −t) φ(x, t, ξ, τ ) = √ . 2πν(τ − t)
(1.1.26)
Notice that the effective range of integration with respect to time in (1.1.22) is 0 < τ < t.
Exercise 1.1.4 (1) Is quantum mechanics captured mathematically as wellposed problems? See, for example, Schiff (1949, p. 48). (2) Can wellposed problems have chaotic solutions?
1.2
Inverse models
1.2.1
Overdetermined problems
We shall spoil the wellposedness of the forward model examined in §1.1, by introducing additional information about the toy “ocean circulation” ﬁeld u(x, t). This information will consist of imperfect observations of u at isolated points in space and time, for the sake of simplicity. The forward model becomes overdetermined; it cannot be solved with smooth functions and must be regarded as illposed. We shall resolve the illposed problem by constructing a weighted, leastsquares bestﬁt to all the information. It will be shown that this bestﬁt or “generalized inverse” of the illposed problem obeys the Euler–Lagrange equations. 1.2.2
Toy ocean data
Let us assume that a ﬁnite number M of measurements (observations, data, . . .) of u were collected in the bounded “ocean basin” 0 ≤ x ≤ L, during the “cruise” 0 ≤ t ≤ T . The data were collected at the points (xm , tm ), where 1 ≤ m ≤ M: see Fig. 1.2.1. The
1.2 Inverse models
13
t
Figure 1.2.1 Toy ocean data.
T (xM, tM)
(xm, tm) (x4, t4) (x3, t3) (x1, t1)
(x2, t2)
0
x L
data are related to the “true” ocean circulation ﬁeld u(x, t) by dm = u(xm , tm ) + m ,
(1.2.1)
1 ≤ m ≤ M, where dm is the datum or recorded value, and u(xm , tm ) is the true value of the circulation. The measurement error m may arise from an imperfect measuring system, or else from mistakenly identifying streamfunction and pressure, for example. On the other hand, if our ocean model were quasigeostrophic and the data included internal waves, then there would also be cause to admit errors in the dynamics.
1.2.3
Failure of the forward solution
Let us now consider how these data relate to the forward problem. If u F = u F (x, t) is the forward solution: ∂u F ∂u F +c =F ∂t ∂x
(1.2.2)
for 0 ≤ x ≤ L and 0 ≤ t ≤ T , with u F (x, 0) = I (x)
(1.2.3)
u F (0, t) = B(t)
(1.2.4)
for 0 ≤ x ≤ L, and
for 0 ≤ t ≤ T , then we may expect that u F (xm , tm ) = dm
(1.2.5)
for at least some m: 1 ≤ m ≤ M. We therefore assume that there are errors in our prior estimates for F, I and B. So the true circulation u must satisfy ∂u ∂u +c =F+ f ∂t ∂x
(1.2.6)
1. Variational assimilation
14
for 0 ≤ x ≤ L and 0 ≤ t ≤ T , u(x, 0) = I (x) + i(x)
(1.2.7)
u(0, t) = B(t) + b(t)
(1.2.8)
for 0 ≤ x ≤ L and
for 0 ≤ t ≤ T . Note that what is implied to be a forcing error f = f (x, t) on the rhs of (1.2.6) may actually be an error in the dynamics expressed on the lefthand side (lhs) of (1.2.6).
1.2.4
Leastsquares ﬁtting: the penalty functional
We have established that for any choice of F + f , I + i and B + b, there is a unique solution for u. However, we have only the M data values dm to guide us and so the error ﬁelds f , i and b are undetermined, while the data errors m are unknown. We ˆ shall seek the ﬁeld uˆ = u(x, t) that corresponds to the smallest values for f , i, b and m in a weighted, leastsquares sense. Speciﬁcally, we shall seek the minimum of the quadratic penalty functional or cost functional J : T J = J [u] ≡ W f
L d x f (x, t) + Wi
dt 0
L
0
T d x i(x) + Wb
2
dt b(t)2 + w
2
0
0
M
m2 ,
m=1
(1.2.9) where W f , Wi , Wb and w are positive weights that we are free to choose. There are more general quadratic forms, but (1.2.9) will sufﬁce for now. The lhs of (1.2.9) expresses the dependence of J upon u, while the rhs only involves f , i, b and m . It is to be understood that the latter are the values that would be obtained, were u substituted into (1.2.1) and (1.2.6)–(1.2.8). These deﬁnitions could be appended to J using Lagrange multipliers, but it is simpler just to remember them ourselves. Finally, note that while u is a ﬁeld of values for 0 ≤ x ≤ L and 0 ≤ t ≤ T , the penalty functional J [u] is a single number for each choice of the entire ﬁeld u. Rewriting (1.2.9), with f , i, b and m replaced by their deﬁnitions, yields T J [u] = W f
L dt
0
dx 0
∂u ∂u +c −F ∂t ∂x
T + Wb
dt {u(0, t) − B(t)}2 + w 0
2
L + Wi
d x {u(x, 0) − I (x)}2 0
M
{u(xm , tm ) − dm }2 .
m=1
The dependence upon u (and upon F, I, B and dm ) is now explicit.
(1.2.10)
1.2 Inverse models
1.2.5
15
The calculus of variations: the Euler–Lagrange equations
We shall use the calculus of variations (Courant and Hilbert, 1953; Lanczos, 1966) to ﬁnd a local extremum of J . Since J is quadratic in u and clearly nonnegative, the ˆ local extremum must be the global minimum. To begin, let uˆ = u(x, t) be the local extremum. That is, ˆ + O(δu)2 J [uˆ + δu] = J [u]
(1.2.11)
for some small change δu = δu(x, t). This statement can be made more precise but we shall proceed informally: ˆ δJ ≡ J [uˆ + δu] − J [u] L T ∂ uˆ ∂δu ∂δu ∂ uˆ = 2W f +c −F +c dt dx ∂t ∂x ∂t ∂x 0
0
L
T ˆ d x {u(x, 0) − I (x)} δu(x, 0) + 2Wb
+ 2Wi 0
+ 2w
M
ˆ t) − B(t)}δu(0, t) dt {u(0, 0
ˆ m , tm ) − dm }δu(xm , tm ) + O(δu)2 . {u(x
(1.2.12)
m=1
Note 1. F, I , B and dm have not been allowed to vary; only uˆ has been varied. Note 2. We have assumed that ∂u ∂δu δ (x, t) = (x, t), etc. (1.2.13) ∂t ∂t The lhs of (1.2.13) is a variation of (∂u/∂t); the rhs is the time derivative of the variation of u. For convenience let us introduce the ﬁeld λ = λ(x, t): ∂ uˆ ∂ uˆ λ ≡ Wf +c −F . ∂t ∂x
(1.2.14)
Then the ﬁrst term in δJ is T 2
dx λ
dt 0
L 0
∂δu ∂δu +c ∂t ∂x
L t=T T x=L = 2 d x λδu + 2 dt λcδu 0
+2
L dt
0
0
t=0
T
dx 0
−
∂λ ∂λ −c ∂t ∂x
x=0
δu(x, t).
(1.2.15)
16
1. Variational assimilation
Notice that the last explicit term in δJ may be written as T 2
L
0
M
dx w
dt
ˆ m , tm ) − dm }δu(x, t)δ(x − xm )δ(t − tm ), {u(x
(1.2.16)
m=1
0
where the second and third δs denote Dirac delta functions. We have now expressed δJ entirely in terms of δu(x, t). None of δu t , δu x and δu(xm , tm ) still appear. We now argue that δJ = O(δu)2 ,
(1.2.17)
implying that uˆ is an extremum of J , provided that the coefﬁcients of δu(x, t), δu(L , t), δu(0, t), δu(0, x) and δu(T, x) all vanish. Examination of (1.2.12), (1.2.15) and (1.2.16) shows that these conditions are, respectively, −
M ∂λ ∂λ −c +w {uˆ m − dm }δ(x − xm )δ(t − tm ) = 0, ∂t ∂x m=1
(1.2.18)
λ(L , t) = 0,
(1.2.19)
ˆ t) − B(t)} = 0, −cλ(0, t) + Wb {u(0,
(1.2.20)
ˆ −λ(x, 0) + Wi {u(x, 0) − I (x)} = 0,
(1.2.21)
λ(x, T ) = 0,
(1.2.22)
ˆ m , tm ). Recall the deﬁnition of λ: where uˆ m ≡ u(x ∂ uˆ ∂ uˆ λ ≡ Wf +c −F . ∂t ∂x
(1.2.23)
These conditions (1.2.18)–(1.2.23) constitute the Euler–Lagrange equations for local extrema of the penalty functional J deﬁned in (1.2.10). How shall we untangle them, to ﬁnd our bestﬁt estimate uˆ of the ocean circulation u? Note 1. Substituting (1.2.23) into (1.2.18) yields “the” Euler–Lagrange equation familiar to physicists. Note 2. Students sometimes derive (1.2.12) from (1.2.10) by expanding the squares ˆ and then subtracting. It is less in the integrand, evaluating at uˆ + δu and at u, tedious to calculate as follows: δJ [u]
T u=uˆ
= δW f
dx
0
0
T
L
= Wf
0
∂u ∂u +c −F ∂t ∂x
dx δ
dt 0
L dt
2
∂u ∂u +c −F ∂t ∂x
+ ··· 2 + ···
1.2 Inverse models
T = Wf
L dt
0
dx 2 0
T = 2W f
dt 0
L dx 0
∂ uˆ ∂ uˆ +c −F ∂t ∂x
∂ uˆ ∂ uˆ +c −F ∂t ∂x
17
∂u ∂u δ +c − F + ··· ∂t ∂x
∂δu ∂δu +c ∂t ∂x
+ ···, (1.2.24)
as in (1.2.12).
Exercise 1.2.1 (requires care) Consider the integral T I=
L d xλ2 .
dt 0
(1.2.25)
0
Substitute for one of the factors of λ in (1.2.25), using (1.2.23). Integrate by parts, and use (1.2.18)–(1.2.22). Conclude that if W f , Wb , Wi and w > 0, then the Euler– Lagrange equations (1.2.18)–(1.2.23) have a unique solution. Discuss the case Wi = 0; it occurs widely in the published literature (Bennett and Miller, 1991).
Exercise 1.2.2 Consider slow, viscous ﬂow driven by an externally imposed pressure gradient, as in Exercise 1.1.3. Assume measurements of u are available, as §1.2.2. Resolve this illposed problem by deﬁning a generalized inverse in terms of a weighted, leastsquares best ﬁt to all the information. Derive the Euler–Lagrange equations, and prove that they have at most one solution.
Exercise 1.2.3 Introduce a forcing error t f nk into the ﬁnitedifference model (1.1.18). By analogy to (1.2.9), a simple penalty function is 2 J [u] = W f (1.2.26) f nk xt + · · · , k
n
where the ellipsis indicates initial penalties, etc., that will be considered below in stages, as will the ranges of the summations in (1.2.26). (i) Show that the Euler–Lagrange equation for extrema of J with respect to variations of u kn is λk−1 − λkn − c(t/x) λkn − λkn+1 = · · · , (1.2.27) n where the ellipses indicate contributions from variations of data penalties in (1.2.26).
18
1. Variational assimilation
(ii) The range of summation over the time index k in (1.2.26) is 0 ≤ k ≤ K − 1, where K t = T . By analogy to (1.2.9), a simple initial penalty is J [u] = · · · + Wi
u 0n − In
2
+ ···.
(1.2.28)
n
Show that, for extrema with respect to u 0n and u nK , −λ0n + Wi uˆ 0n − In = 0
(1.2.29)
λnK −1 = 0,
(1.2.30)
and
respectively. Compare these with (1.2.21) and (1.2.22). (iii) Choose a range of summation over the space index n in (1.2.26), and prescribe a simple boundary penalty analogous to that in (1.2.9). Derive extremal conditions analogous to (1.2.19) and (1.2.20). (iv) Assume that there are M measurements of u kn , that is, measured values of u(x, t) at grid points in space and time. Prescribe a simple data penalty as in (1.2.9) and derive the contributions to (1.2.27) from variations of this data penalty. Hint: Replace the Dirac delta functions δ(xn − xm ) and δ(tk − t j ) with (x)−1 δnm and (t)−1 δk j respectively, where δnm is the Kronecker delta: δnm =
1 0
n=m n = m.
(1.2.31)
1.3
Solving the Euler–Lagrange equations using representers
1.3.1
Leastsquares ﬁtting by explicit solution of extremal conditions
The mixed initialboundary value problem (1.1.1)–(1.1.3) for the ﬁrstorder wave equation, together with the data (1.2.1) and the simple leastsquares penalty functional (1.2.10), have led us to the awkward system of Euler–Lagrange equations (1.2.18)– ˆ It may be obtained explicitly, by (1.2.23). The solution is the bestﬁt ocean circulation u. an intricate construction involving “representer functions”. The effort is rewarded not only with structural insight, but also with enormous gains in computational efﬁciency compared to conventional minimization of (1.2.10) using gradient information.
1.3 Representers
1.3.2
19
The Euler–Lagrange equations are a twopoint boundary value problem in time
After a little reordering, the Euler–Lagrange equations for local extrema uˆ of the penalty functional J [u] are: M ∂λ ∂λ {uˆ m − dm }δ(x − xm )δ(t − tm ) − ∂t − c ∂ x = −w (1.3.1) m=1 (B) λ(x, T ) = 0 (1.3.2) λ(L , t) = 0, (1.3.3) ∂ uˆ + c ∂ uˆ = F + W −1 λ f ∂x ∂t −1 (F) u(x, ˆ 0) = I (x) + Wi λ(x, 0) ˆ t) = B(t) + cWb−1 λ(0, t). u(0,
(1.3.4) (1.3.5) (1.3.6)
Note 1. Our best estimates for f, i and b are fˆ(x, t) ≡ W −1 f λ(x, t),
ıˆ(x) ≡ Wi−1 λ(x, 0),
ˆ ≡ cWb−1 λ(0, t). b(t) (1.3.7)
Note 2. Eq. (1.3.1) is known as the “backward” or “adjoint” equation. Note 3. At ﬁrst glance, it would seem that we could proceed by integrating the ˆ t), system (B) “backwards in time and to the left” (see Fig. 1.2.1), yielding λ(x, ˆλ(0, t) and λ(x, ˆ 0). Then we could integrate the system (F) “forwards and to the ˆ right” (see Fig. 1.2.1), yielding the ocean circulation estimate uˆ = u(x, t). However, after reexamining (1.3.1), we see that it is necessary to know ˆ m , tm ) in order to integrate (B). The Euler–Lagrange equations do not u(x consist of two initialvalue problems; they constitute a single, twopoint boundary value problem in the time interval 0 ≤ t ≤ T .
1.3.3
Representer functions: the explicit solution and the reproducing kernel
Let us introduce the representer functions. There are M of them, denoted by rm (x, t), 1 ≤ m ≤ M. Each has an “adjoint” αm (x, t), satisfying ∂α ∂αm m − ∂t − c ∂ x = δ(x − xm )δ(t − tm ) (Bm ) αm (x, T ) = 0 αm (L , t) = 0.
(1.3.8) (1.3.9) (1.3.10)
As a consequence of the single impulse on the rhs of (1.3.8) being “bare”, we may integrate (Bm ) “backwards and to the left”, yielding αm (x, t). We may then solve for
20
1. Variational assimilation
rm by integrating (Fm ) “forward and to the right”: ∂rm ∂rm −1 ∂t + c ∂ x = W f αm (Fm ) rm (x, 0) = Wi−1 αm (x, 0) rm (0, t) = cWb−1 αm (0, t).
(1.3.11) (1.3.12) (1.3.13)
Next, we seek a solution of (1.3.1)–(1.3.6) in the form ˆ u(x, t) = u F (x, t) +
M
βm rm (x, t),
(1.3.14)
m=1
where u F is the prior estimate (the solution of the forward model (1.2.2)–(1.2.4)), and the βm are unknown constants. If we substitute (1.3.14) into (1.3.4), and derive D uˆ = Du F +
M
βm Drm
(1.3.15)
m=1
= F + W −1 f
M
βm αm ,
(1.3.16)
m=1
where D =
∂ ∂t
+ c ∂∂x , we ﬁnd that λ ≡ W f {D uˆ − F} =
M
β m αm .
(1.3.17)
m=1
Furthermore, −Dλ = −
M
βm Dαm
m=1
=
M
βm δ(x − xm )δ(t − tm )
(1.3.18)
m=1
= −w
M
{uˆ m − dm }δ(x − xm )δ(t − tm ),
(1.3.19)
m=1
by virtue of (1.3.1). Equating coefﬁcients of the impulses, we obtain the optimal choices βˆ m for the representer coefﬁcients βm : βm = βˆ m ≡ −w{uˆ m − dm } for 1 ≤ m ≤ M. Substituting again for uˆ m yields M βˆ m = −w u Fm + βˆ l rlm − dm l=1
where u Fm ≡ u F (xm , tm ) and rlm ≡ rl (xm , tm ).
(1.3.20)
,
(1.3.21)
1.3 Representers
21
Hence M
(rlm + w −1 δlm )βˆ l = h m ≡ dm − u Fm ,
(1.3.22)
l=1
where δlm is the Kronecker delta. In matrix notation, the M equations (1.3.22) for the M representer coefﬁcients βˆ m become (R + w −1 I)βˆ = h ≡ d − u F .
(1.3.23)
Note 1. The rhs h is known; it is the data vector minus the vector of measured values of the prior estimate. Note 2. The diagonal weight matrix wI is readily generalized to symmetric positive deﬁnite matrices w. Note 3. The l th column of the M × M “representer matrix” R consists of the M measured values of the l th representer function rl (x, t). Note 4. It will be shown (see (1.3.32)) that R is symmetric: R = RT . ˆ Note 5. The generalized inverse problem of ﬁnding the ﬁeld uˆ = u(x, t), where 0 ≤ x ≤ L and 0 ≤ t ≤ T , has been exactly reduced to the problem of inverting ˆ an M × M matrix, in order to ﬁnd the M representer coefﬁcients β. ˆ Finally, we have an explicit solution for u: ˆ u(x, t) = u F (x, t) + (d − u F )T (R + w−1 )−1 r(x, t).
(1.3.24)
It was established in §1.1 that the forward model (1.1.1)–(1.1.3) has a unique solution for each choice of the inputs. Accordingly, the partial differential operator in (1.1.1), the initial operator in (1.1.2) and the boundary operator in (1.1.3) constitute a nonsingular operator. It may be inverted; the inverse operator is expressed explicitly in (1.1.17) with the Green’s function γ . Introducing the measurement operators as in (1.2.1) yields a problem with no solution, thus the operator comprising those in (1.1.1)–(1.1.3) and (1.2.1) is singular; it is not invertible in the regular sense. However, a generalized inverse has been deﬁned in the weighted leastsquares sense of (1.2.10), and is explicitly expressed in (1.3.24) with the representers for the penalty functional (1.2.10), and with the Green’s function for the nonsingular operator. Recall that u F is given by (1.1.17), although it will in practice be computed by numerical integration of (1.2.2)–(1.2.4). In an abuse of language, we shall refer to the bestﬁt uˆ given by (1.3.24) as the generalized inverse estimate, or simply the inverse.
Exercise 1.3.1 Verify that the initial condition (1.3.5) and boundary conditions (1.3.6) are satisﬁed.
22
1. Variational assimilation
In summary, the steps for solving the Euler–Lagrange equations are: (1) (2) (3) (4)
calculate u F (x, t) and hence u F ; calculate r(x, t) and hence R; invert P ≡ R + w−1 ; assemble (1.3.24).
Note 1. u F (x, t) depends upon the “dynamics”, the initial operator, the boundary operator and the choices for F, I and B. M Note 2. u F depends upon u F and the “observing network” {(xm , tm )}m=1 . Note 3. r depends upon the dynamics, the initial operator, the boundary operator, the −1 −1 observing network and the inverted weights W −1 f , Wi , Wb . Note 4. βˆ depends upon R, the inverse of the data weight w, and the prior data misﬁt h ≡ d − uF . Note 5. See Fig. 3.1.1 for a “time chart” implementing the representer solution.
Exercise 1.3.2 Express λ, fˆ, ıˆ and bˆ using representer functions and their adjoints.
Exercise 1.3.3 (trivial) Show that J F ≡ J [u F ] = hT w h.
(1.3.25)
Exercise 1.3.4 (nontrivial) Show that (i) (ii)
ˆ = hT P−1 h, Jˆ ≡ J [u] ˆ ˆ T w(d − u) ˆ = hT P−1 w−1 P−1 h, Jdata ≡ (d − u)
(1.3.26) (1.3.27)
and (iii)
Jˆmod ≡ Jˆ − Jˆdata = hT P−1 RP−1 h.
Note that Jˆmod is the sum of dynamical, initial and boundary penalties.
(1.3.28)
Let us now prove that the representer matrix is symmetric: R = RT . First, recall that the adjoint representer αm (x, t) for a point measurement at (xm , tm ) is just the Green’s function γ (x, t, xm , tm ), where γ (x, t, y, s) satisﬁes −
∂γ ∂γ −c = δ(x − y)δ(t − s), ∂t ∂x
(1.3.29)
1.4 Limiting weights
23
subject to γ = 0 at t = T , γ = 0 at x = L. Now let (x, t, y, s) satisfy ∂ ∂ + c = W −1 f γ, ∂t ∂x
(1.3.30)
subject to = Wi−1 γ at t = 0, and = cWb−1 γ at x = 0. Thus rm (x, t) = (x, t, xm , tm ).
Exercise 1.3.5 (Bennett, 1992) Show that
dz dr γ (z, r, x, t)γ (z, r, y, s) −1 + Wi dz γ (z, 0, x, t)γ (z, 0, y, s) + c2 Wb−1 dr γ (0, r, x, t)γ (0, r, y, s).
(x, t, y, s) = W −1 f
(1.3.31)
Hence representers are not Green’s functions; rather they are “squares” of Green’s functions. Note that is symmetric, but γ is not symmetric. Finally we deduce that rlm ≡ rl (xm , tm ) = (xm , tm , xl , tl ) = (xl , tl , xm , tm ) = rm (xl , tl ) ≡ rml .
(1.3.32)
That is, R = RT . Note that is known as a “reproducing kernel” or “rk”, for reasons given in §2.1.
1.4
Some limiting choices of weights: “weak” and “strong” constraints
1.4.1
Diagonal data weight matrices, for simplicity
The parade of formulae in the previous sections should become more meaningful as we explore some limiting choices for the weights. We shall assume that the data weight matrix is diagonal: w = wI,
(1.4.1)
in order to avoid technicalities such as the norm of a matrix. Note that (1.4.1) implies w−1 = w −1 I.
(1.4.2)
24
1.4.2
1. Variational assimilation
Perfect data
If we believe that the data are perfectly accurate, then we should give inﬁnite weight to them. In this case we hope that the inverse estimates agree exactly with the data, at the measurement sites. Let us therefore consider the limit: w → ∞. Hence P ≡ R + w −1 I → R,
(1.4.3)
βˆ → R−1 h,
(1.4.4)
ˆ u(x, t) → u F (x, t) + r(x, t)T R−1 h.
(1.4.5)
and
Measuring both sides of (1.4.5) yields uˆ → u F + RT R−1 h,
(1.4.6)
= uF + h
(1.4.7)
= u F + (d − u F )
(1.4.8)
= d,
(1.4.9)
as required. Note that we have used the symmetry of the representer matrix: RT = R. In this limit, the inverse estimate interpolates the data. 1.4.3
Worthless data
Now suppose that we believe the data are worthless, that is, we have no information about the magnitude of the data errors. In practice we always have some idea: the errors in altimetry data do not exceed the height of the orbit of the satellite, but that is inﬁnite by any hydrographic standard. We should therefore consider the limit: w → 0. Then P−1 = (R + w −1 I)−1 → 0,
(1.4.10)
ˆ u(x, t) → u F (x, t).
(1.4.11)
hence
That is, the data have no inﬂuence on the inverse estimate, as would be desirable. 1.4.4
Rescaling the penalty functional
The Euler–Lagrange equations for local extrema of J [u] are also those for local extrema of 2J [u]. This is true even if the dynamics and observing systems are nonlinear, or if J is not quadratic. Thus the limiting cases: w → ∞, w → 0 really refer to w/W f , w/Wi , w/Wb all → ∞, or all → 0. That is, they refer to the relative weighting of the various information. However, the prior and posterior functional values J F ≡ J [u F ], ˆ do depend upon the absolute values of the weights. This will be crucial and Jˆ ≡ J [u]
1.4 Limiting weights
25
later, when we interpret these numbers as test statistics for the model as a formal hypothesis about the ocean.
1.4.5
Perfect dynamics: Lagrange multipliers for strong constraints
The admission of the error ﬁeld f = f (x, t) in (1.2.6), and the inclusion of !L !T W f 0 d x 0 dt f 2 in the penalty functional (1.2.9), leads to the model being described as a “weak constraint” upon the inversion process (Sasaki, 1970). The model may alternatively be imposed as a “strong constraint”. In that case, the penalty functional is L K[u] = Wi
d x i(x)2 + Wb
dt b(t)2 + w
M
m2 .
(1.4.12)
m=1
0
Compare (1.4.12) and (1.2.9): i, b and m are deﬁned as before by (1.2.7), (1.2.8) and (1.2.1) respectively, but now we require that u = u(x, t) satisfy (1.1.1) exactly. This requirement may be met in the search for the minimum of K, by appending the strong constraint (1.1.1) to K using a Lagrange multiplier ﬁeld ψ = ψ(x, t): L L[u, ψ] = K[u] + 2
T dt ψ(x, t)
dx 0
0
∂u ∂u (x, t) + c (x, t) − F(x, t) . ∂t ∂x (1.4.13)
The factor of two will be seen to be convenient. Note that the augmented penalty functional L depends on u and ψ, which may vary independently. The total variation in L is L δL = δK + 2 0
+2
dt δψ
dx
L
0
T dx
0
T
0
∂u ∂u +c −F ∂t ∂x
∂u ∂u dt ψ δ + cδ ∂t ∂x
+ O(δ 2 ).
(1.4.14)
ˆ ˆ If the pair of ﬁelds ψˆ = ψ(x, t) and uˆ = u(x, t) extremize L for arbitrary variations δψ and δu, then ∂ψ ∂ψ {uˆ m − dm } δ(x − xm )δ(t − tm ) −c = −w ∂t ∂x m=1 M
−
ψ(x, T ) = 0
(1.4.16)
ψ(L , t) = 0 ˆ ∂u ∂ uˆ +c =F ∂t ∂x ˆ u(x, 0) = I (x) + Wi−1 ψ(x, 0) ˆ t) = B(t) + u(0,
(1.4.15)
cWb−1 ψ(0, t).
(1.4.17) (1.4.18) (1.4.19) (1.4.20)
26
1. Variational assimilation
The strong constraint (1.4.18) is immediately recovered from (1.4.14), if δL = 0 and δψ is arbitrary. The other Euler–Lagrange conditions are recovered as in §1.2. Comparing (1.4.15)–(1.4.20) with (1.3.1)–(1.3.6) establishes that ψ(x, t) = lim λ(x, t) . W f →∞
(1.4.21)
It would seem from (1.3.1)–(1.3.3) that λ is independent of W f , but there is an implicit dependence through uˆ m , 1 ≤ m ≤ M, in (1.3.1). Thus, we may recover the “strong constraint” inverse from the “weak constraint” inverse in the limit as W f → ∞.
1.5
Regularity of the inverse estimate
1.5.1
Physical realizability
Thus far our construction has been formal: we paid no attention to the physical realizˆ We shall now see that uˆ is in fact unrealistic, unless ability of the inverse estimate u. we make more interesting choices for the weights W f , Wi and Wb .
1.5.2
Regularity of the Green’s functions and the adjoint representer functions
Consider “the” Euler–Lagrange equation: −
M ∂λ ∂λ (dm − uˆ m )δ(x − xm )δ(t − tm ). −c =w ∂t ∂x m=1
(1.5.1)
In fact, just consider the equation for an adjoint representer function: −
∂αm ∂αm −c = δ(x − xm )δ(t − tm ), ∂t ∂x
(1.5.2)
subject to αm (x, T ) = 0,
αm (L , t) = 0 .
(1.5.3)
The solution is the Green’s function: αm (x, t) = γ (x, t, xm , tm ) = δ(x − xm − c(t − tm ))H (tm − t),
(1.5.4)
and λ(x, t) = β T α(x, t). Clearly the αm and hence λ are singular, and not just at the data points (xm , tm ): see Fig. 1.5.1. Now uˆ obeys ∂ uˆ ∂ uˆ +c = F + fˆ = F + W −1 f λ, ∂t ∂x
(1.5.5)
1.5 Regularity of the inverse estimate
27
t
Figure 1.5.1 Support of αm (x, t). The arrows (the delta functions) are normal to the page.
T
(xm, tm)
tmc 1xm
x 0
L
t
Figure 1.5.2 Support of ˆ singularities in u(x, t).
T (xM, tM)
(xm, tm) (x4, t4) (x3, t3) (x1, t1)
(x2, t2)
0
x L
subject to the initial and boundary conditions uˆ = I + ıˆ = I + Wi−1 λ,
uˆ = B + bˆ = B + cWb−1 λ .
(1.5.6)
So our estimates of fˆ, ıˆ and bˆ are singular. There is neither dispersion nor diffusion in our “toy” ocean dynamics, so uˆ is also singular: see Fig. 1.5.2. This is hardly a satisfactory combination of dynamics and data!
Exercise 1.5.1 Express rm and uˆ using the Green’s function γ . 1.5.3
Nondiagonal weighting: kernel inverses of weights
We want the data to inﬂuence the circulation at remote places and times, so we should give weight to products of residuals at remote places and times. We therefore generalize
28
1. Variational assimilation
the penalty functional (1.2.9) to T J [u] =
T dt
0
L ds
0
L +
dx 0
dy f (x, t)W f (x, t, y, s) f (y, s) 0
L
T dy i(x)Wi (x, y)i(y) +
dx 0
+
L
0
dt 0
M M
T ds b(t)Wb (t, s)b(s) 0
l wlm m .
(1.5.7)
l=1 m=1
Thus our previous, trivial choices were W f (x, t, y, s) = W f · δ(x − y)δ(t − s),
etc.
(1.5.8)
The notations T •≡
L dt
0
L ◦≡
d x, 0
T d x,
∗≡
0
dt 0
allow us to write J more compactly as J [u] = f • W f • f + i ◦ Wi ◦ i + b ∗ Wb ∗ b + T w.
(1.5.9)
Exercise 1.5.2 Deﬁne the weighted residual or adjoint variable λ(x, t) by ∂ uˆ ∂ uˆ λ ≡ Wf • +c −F . ∂t ∂x
(1.5.10)
Then show that the Euler–Lagrange equations for minima of (1.5.9) are just as before.
Exercise 1.5.3 Deﬁne C f , the inverse of W f , by T Cf • Wf ≡
L dr
0
dz C f (x, t, z, r )W f (z, r, y, s)
(1.5.11)
0
= δ(x − y)δ(t − s) .
(1.5.12)
Deﬁne Ci and Cb analogously, and deﬁne C by wC = I .
(1.5.13)
1.5 Regularity of the inverse estimate
29
Each entity in (1.5.13) is an M × M matrix. Now, write out the representer solution of the Euler–Lagrange equations. Verify that the solution only requires C f , Ci , Cb and C ; that is, it does not require their inverses, the weights W f , Wi , Wb and w.
1.5.4
The inverse weights smooth the residuals
The inverse estimate uˆ obeys ∂ uˆ ∂ uˆ (x, t) + c (x, t) = F(x, t) + (C f • λ)(x, t) ∂t ∂x T L = F(x, t) + ds dy C f (x, t, y, s)λ(y, s), 0
(1.5.14)
0
subject to ˆ u(x, 0) = I (x) + (Ci ◦ λ)(x, 0),
(1.5.15)
ˆ t) = B(t) + c(Cb ∗ λ)(0, t) . u(0,
(1.5.16)
and
The supposition is that C f , Ci and Cb smooth the singular behavior of λ, yielding regular estimates for fˆ ≡ C f • λ, ıˆ ≡ Ci ◦ λ and bˆ = cCb ∗ λ, leading in turn to a regular estimate uˆ for the ocean circulation. In summary, we should avoid “diagonal” weighting. Note 1. The adjoint variables α and λ remain singular, but r and u should become regular. Note 2. Evaluation of the convolutions in (1.5.14), (1.5.15) and (1.5.16) at each position and time is potentially very expensive: consider three space dimensions and time. Note 3. Functional analysis sheds much light on smoothness: see §2.6.
Exercise 1.5.4 Consider slow, viscous ﬂow as discussed in Exercises 1.1.3 and 1.2.2. Construct both the adjoint representers α and the representers r, using the Green’s function φ given in (1.1.26). How smooth are α and r? Is nondiagonal weighting of either the dynamical penalty or the initial penalty necessary?
Exercise 1.5.5 Generalize the deﬁnition of the rk given in (1.3.30) et seq. Prove that
30
(i) (ii)
1. Variational assimilation
rm (x, t) = (x, t, xm , tm ), for 1 ≤ m ≤ M; = γ • C f • γ + γ ◦ Ci ◦ γ + c2 γ ∗ C b ∗ γ .
(1.5.17) (1.5.18)
Note: The adjoint equations (1.3.1)–(1.3.3) and forward equations (1.5.14)–(1.5.16), which constitute the most general form of the Euler–Lagrange equations developed in §1.2 and §1.5, are restated for convenience in §4.2 as (4.2.1)–(4.2.6).
Chapter 2 Interpretation
The calculus of variations uses Green’s functions and representers to express the best ﬁt to a linear model and data. Mathematical construction of the representers is devious, and the meaning of the representer solution to the “control problem” of Chapter 1 is not obvious. There is a geometrical interpretation, in terms of observable and unobservable degrees of freedom. Unobservability deﬁnes an orthogonality, and the representers span a ﬁnitedimensional subspace of the space of all model solutions or “circulations”. The representers are in fact the observable degrees of freedom. A statistical interpretation is also available: if the unknown errors in the model are regarded as random ﬁelds having prescribed means and covariances, then the representers are related, via the measurement processes, to the covariances of the circulations. Thus the representer solution to the variational problem is also the optimal linear interpolation, in time and space, of data from multivariate, inhomogeneous and nonstationary random ﬁelds. The minimal value of the penalty functional that deﬁnes the generalized inverse or control problem is a random number. It is the χ 2 variable, if the prescribed error means or covariances are correct, and has one degree of freedom per datum. Measurements need not be pointwise values of the circulation; representers along with their geometrical and statistical interpretations may be constructed for all bounded linear measurement functionals. Analysis of the conditioning of the determination of the representer amplitudes reveals those degrees of freedom which are the most stable with respect to the observations. This characterization also indicates the efﬁciency of the observing system – the fewer unstable degrees of freedom, the better. Interpreting the variational formulation is completed by demonstrating the relationship between weights, covariances and roughness penalties.
31
32
2. Interpretation
2.1
Geometrical interpretation
2.1.1
Alternatives to the calculus of variations
After formulating the penalty functional that deﬁnes the best ﬁt to our model and our data, we found a local extremum using the theory of the calculus of the ﬁrst variation. Speciﬁcally, we derived the Euler–Lagrange equations, and explicitly expressed their solution with representers. These functions were deﬁned as special and directly calculable solutions of Euler–Lagrangelike equations. We shall now construct the same extremum for the penalty functional using Hilbert Space theory (Yoshida, 1980). This geometrical construction reveals the efﬁciency of minimization algorithms based on the Euler–Lagrange equations.
Exercise 2.1.1 How do we know that we shall ﬁnd the same extremum?
2.1.2
Inner products
We begin by deﬁning an inner product for two “ocean circulations” u = u(x, t) and v = v(x, t): ∂ ∂ +c u(x, t) ∂t ∂x 0 0 0 0 ∂ ∂ ×W f (x, t, y, s) +c v(y, s) ∂s ∂y L L + d x dy u(x, 0)Wi (x, y)v(y, 0)
T u, v ≡
T
dt
L
ds
0
0
T
T
+
dt 0
L
dx
dy
ds u(0, t)Wb (t, s)v(0, s) 0
= f u • W f • f v + u ◦ Wi ◦ v + u ∗ Wb ∗ v, where f u is the residual for u (Bennett, 1992).
Exercise 2.1.2 Verify that , is an inner product, that is: (i) (ii) (iii) (iv)
u, v = v, u (assume the W are symmetric), cu + dw, v = cu, v + dw, v for all real numbers c and d, u, u ≥ 0, u, u = 0 ⇔ u ≡ 0 (nontrivial).
(2.1.1)
2.1 Geometrical interpretation
33
In terms of the inner product, our penalty functional is J [u] = u − u F , u − u F + (d − u)T w(d − u).
2.1.3
(2.1.2)
Linear functionals and their representers; unobservables
Consider the linear mapping u → u(ξ, τ ),
(2.1.3)
where the lhs is a ﬁeld, while the rhs is a particular value of the ﬁeld. This mapping is a linear functional: it linearly maps a function to a single number.
Theorem 2.1.1 If the vector space of admissible ﬁelds u, with the inner product , , is complete (that is, if it is a Hilbert Space), then there is a function ρ(x, t, ξ, τ ) such that ρ, u = u(ξ, τ ).
(2.1.4)
So ρ “represents” the measurement process. This is the Riesz representation theorem. Given ρ, we may express J entirely in terms of inner products (Wahba and Wendelberger, 1980): J [u] = u − u F , u − u F + (d − ρ, u)T w(d − ρ, u),
(2.1.5)
where ρm = ρ(x, t, xm , tm ), 1 ≤ m ≤ M. Now, any ﬁeld u = u(x, t) may be expressed as u(x, t) = u F (x, t) +
M
νm ρ(x, t, xm , tm ) + g(x, t),
(2.1.6)
m=1
where u F is again the solution of (1.2.2)–(1.2.4), and where the νm are any coefﬁcients, since we may always choose g ≡ u − u F − ν T ρ.
(2.1.7)
Let us now impose the condition that g is “unobservable”: ρm , g = g(xm , tm ) = 0
(2.1.8)
for 1 ≤ m ≤ M. That is, g is orthogonal to each ρm . For a given u and a given u F , we may use (2.1.8) to derive M equations for the νm ; then g is uniquely deﬁned by (2.1.7).
34
2.1.4
2. Interpretation
Geometric minimization with representers
But we’re not given u; we’re only given u F . Thus ν and g are arbitrary. We wish to ﬁnd the u that minimizes J [u]. Let us evaluate J using (2.1.5) and (2.1.6): J [u] = ν T ρ + g, ν T ρ + g + (d − ρ, u F + ρT ν + g)T w(d − ρ, u F + ρT ν + g) = ν T ρ, ρT ν + ν T ρ, g + g, ρT ν + g, g + (d − ρ, u F − ρ, ρT ν − ρ, g)T w(d − ρ, u F − ρ, ρT ν − ρ, g). (2.1.9) Next impose the M orthogonality conditions (2.1.8), and use the representing property of ρ to obtain J [u] = J [ν, g] = ν T ρ, ρT ν + g, g + (d − u F − ρ, ρT ν)T w(d − u F − ρ, ρT ν).
(2.1.10)
The penalty functional J is now expressed explicitly in terms of ν and g. Note that g only appears once on the rhs of (2.1.10). Clearly J is least with respect to the choice of g if g, g = 0, that is g = gˆ ≡ 0.
(2.1.11)
We discard the ﬁeld g orthogonal to all the representers. It remains to select the νm , 1 ≤ m ≤ M. But ﬁrst note that σlm ≡ ρ, ρT lm = ρl , ρm = ρm , ρl = ρm (xl , tl ) = ρl (xm , tm ) = σml ,
(2.1.12)
so σ = σ T and J , which now only depends upon ν, may be expressed as J [u] = J [ν] = ν T σν + (h − σν)T w(h − σν),
(2.1.13)
where h ≡ d − u F . Completing the square, J [ν] = (ν − ν) ˆ T S(ν − ν) ˆ + hT wh − νˆ T Sν, ˆ
(2.1.14)
where S = σ + σwσ and Sνˆ = σwh, both of which are given. We ﬁnally minimize J by choosing ν = ν, ˆ and then Jˆ ≡ J [ν] ˆ = hT wh − νˆ T Sν. ˆ Provided σ is nonsingular, we may untangle these results to ﬁnd (σ + w−1 )νˆ = h,
(2.1.15)
uˆ = u F + νˆ T ρ + gˆ ,
(2.1.16)
which looks familiar. Recall that our minimizer is
where gˆ satisﬁes (2.1.11), and νˆ satisﬁes (2.1.15).
2.1 Geometrical interpretation
2.1.5
35
Equivalence of variational and geometric minimization: the data space
Surely the representers ρ deﬁned by the representing property (2.1.4) are the same as the representer functions r that satisfy the Euler–Lagrangelike system (1.3.8)–(1.3.13), in which case σ = R,
νˆ = βˆ ?
(2.1.17)
ρm (x, t) = rm (x, t)
(2.1.18)
Exercise 2.1.3 Show that
for 0 ≤ x ≤ L, 0 ≤ t ≤ T , and 1 ≤ m ≤ M.
Hint Consider (2.1.4): u(xm , tm ) ≡ ρm , u T T L L ∂ ∂ ≡ dt ds d x dy +c ρm (x, t) ∂t ∂x 0 0 0 0 ∂ ∂ ×W f (x, t, y, s) +c u(y, s) ∂s ∂y L L + dx dy ρm (x, 0)Wi (x, y)u(y, 0) 0
0
T
T
+
ds ρm (0, t)Wb (t, s)u(0, s).
dt 0
(2.1.19)
0
Integrate the ﬁrst integral by parts, and then compare (2.1.19) to T u(xm , tm ) =
L d x u(x, t)δ(x − xm )δ(t − tm ).
dt 0
(2.1.20)
0
Since the partiallyintegrated (2.1.19) must agree with (2.1.20) for all ﬁelds u(x, t) having initial values u(x, 0) and boundary values u(0, t), we may equate their respective coefﬁcients, arriving at (1.3.8)–(1.3.13). We have proved that for any ﬁeld u, r, u = u.
(2.1.21)
36
2. Interpretation
Note 1. We have established that T ˆ u(x, t) − u F (x, t) = βˆ r(x, t).
(2.1.22)
Nu ll
That is, the difference between the inverse estimate uˆ and the prior estimate u F is a linear combination of the M representers r1 , . . . , r M . The difference lies in the observable space, that is, we reject any additional difference g that is unobservable: r, g = 0. We began with a search for the optimal or bestﬁt ˆ ﬁeld u(x, t), where 0 ≤ x ≤ L and 0 ≤ t ≤ T . This would be a search amongst an inﬁnite number of degrees of freedom (the “state space”). We have exactly reduced the task to a search for the M optimal representer coefﬁcients βˆ 1 , . . . , βˆ M : see Fig. 2.1.1. These are the observable degrees of freedom (the “data space”).
State
u
State
Da
ta
Figure 2.1.1 The plane represents the state space. It has an axis u(x, t) for each (x, t) in the intervals 0 ≤ x ≤ L, 0 ≤ t ≤ T . In principle this is an inﬁnite dimensional space. In practice, when we replace continuous intervals and partial differential equations with grids and partial difference equations, the state space usually has a very large but ﬁnite dimension. The contour is deﬁned by a constant value for the penalty functional J [u], and has principal axes RData (for β T r(x, t)) and RNull (for g(x, t)). Note that the representers r1 (x, t), . . . , r M (x, t) are known, and span the data space, so only the unknown β1 , . . . , β M vary in the data space. The unobservable ﬁeld g(x, t) is unknown and variable for 0 ≤ x ≤ L, 0 ≤ t ≤ T . Realizing that gˆ is zero greatly ˆ as we need only search in reduces the size of the search for u, the data space.
2.2 Statistical interpretation
37
Note 2. Recall from §1.3.3 the deﬁnition of as the representer for point measurements. Hence for any ﬁeld u, , u = u,
(2.1.23)
and consequently is known as a “reproducing kernel.”
2.2
Statistical interpretation: the relationship to “optimal interpolation”
2.2.1
Random errors
Viewed as a generalized inverse or as a control problem, the ocean circulation u is estimated by adjusting the forcing f , initial value i, and boundary value b in order to obtain a better ﬁt to the data, given that there are weights or “costs” W f , Wi , Wb and w for these control variables and for the data misﬁt. Alternatively, the ﬁelds f , b and i for 0 ≤ x ≤ L and 0 ≤ t ≤ T , and the data error vector , may be viewed as members of an ensemble of such quantities. That is, they are random. We shall attempt to estimate which f , i, b and were present in the “ocean” in 0 ≤ x ≤ L during our “cruise” for 0 ≤ t ≤ T . We shall recover an interpretation of the “variational assimilation” of §1.2 in terms of the “optimal interpolation” routinely used in meteorology and oceanography.
2.2.2
Null hypotheses
In order to make these estimates, we shall have to make some assumptions about the ensemble. These assumptions compose a null hypothesis H0 : E f (x, t) = Ei(x) = Eb(t) = Em = 0
(2.2.1)
for 0 ≤ x ≤ L, 0 ≤ t ≤ T and 1 ≤ m ≤ M, where E( ) denotes the ensemble average or mean; E( f (x, t) f (y, s)) = C f (x, t, y, s), E(i(x)i(y)) = Ci (x, y), E(b(t)b(s)) = Cb (t, s),
(2.2.2)
and E(T ) = C , while E( f i) = E( f b) = E( f m ) = E(ib) = E(im ) = E(bm ) = 0.
(2.2.3)
38
2. Interpretation
Note 1. The covariances C f , Ci , Cb and C are explicit functional or tabular forms. Note 2. Only ﬁrst and second moments are given in H0 = {(2.2.1), (2.2.2), (2.2.3)}; if the random variables f , i, b and are Gaussian, these moments determine the probability distribution function (pdf). Note 3. The alternative hypothesis is that either (2.2.1), (2.2.2), or (2.2.3) is not true.
2.2.3
The reproducing kernel is a covariance
Let us now deﬁne v = v(x, t) by u(x, t) = u F (x, t) + v(x, t).
(2.2.4)
That is, v is the random error in our prior estimate or forward solution u F . The latter corresponds to our prior estimates F, I and B for the forcing, etc. These estimates are made prior to knowing the data d. Clearly ∂v ∂v +c = f, ∂t ∂x
(2.2.5)
subject to v = i at t = 0, and v = b at x = 0. We may use the Green’s function γ to write v = γ • f + γ ◦ i + cγ ∗ b.
(2.2.6)
Ev = 0.
(2.2.7)
Hence
Exercise 2.2.1 Derive in detail the covariance for v: Cv ≡ E(vv) = γ • E( f f ) • γ + γ ◦ E(ii) ◦ γ + c2 γ ∗ E(bb) ∗ γ = γ • C f • γ + γ ◦ Ci ◦ γ + c2 γ ∗ C b ∗ γ
(2.2.8)
= .
(2.2.9)
That is, the covariance of the errors in the prior estimate is just the reproducing kernel (Weinert, 1982; Bennett, 1992). 2.2.4
“Optimal Interpolation”, or best linear unbiased estimation; equivalence of generalized inversion and OI
We shall now outline the method of “optimal interpolation” (OI) for estimating a ﬁeld u, given a ﬁrstguess u F and data d (Bretherton et al., 1976; Daley, 1991; Thi´ebaux and Pedder, 1987). The ﬁrst guess need not be a model solution.
2.2 Statistical interpretation
39
Suppose the true ﬁeld is u(x, t) = u F (x, t) + q(x, t),
(2.2.10)
d = u + ,
(2.2.11)
Eq(x, t) = Em = 0
(2.2.12)
and suppose as before that
where
for 0 ≤ x ≤ L, 0 ≤ t ≤ T and 1 ≤ m ≤ M; and suppose that E(q(x, t)q(y, s)) = Cq (x, t, y, s), = C , E(T ) and E(q(x, t)) =0
(2.2.13)
for 0 ≤ x, y ≤ L, 0 ≤ t, s ≤ T . Note 1. E( ) denotes an ensemble average, given the prior estimate u F . Note 2. Eu = Eu F + Eq = u F + 0 = u F . Note 3. Ed = Eu + E = Eu F + Eq + E = u F + 0 + 0 = u F . We seek the best linear unbiased estimate of u, that is u L (x, t) = u F (x, t) + (d − u F )T s(x, t),
(2.2.14)
where s1 (x, t), . . . , s M (x, t) are M as yet unchosen nonrandom interpolants. Note 1. u L is linear in u F and d. Note 2. u L is unbiased: Eu L = u F = Eu. Note 3. u L is best if Ee2L (x, t) ≡ E{u(x, t) − u L (x, t)}2
(2.2.15)
is least for each (x, t). Now Ee2L = E{u F + q − u F − (u F + q + − u F )T s}2 = E{q − (q + ) s} T
(2.2.16)
2
= Eq 2 − E{q(q + )T } s − sT E{(q + )q} + sT E{(q + )(q + )T }s
(2.2.17)
for each (x, t). We want ∂ Ee2L =0 ∂sm
(2.2.18)
40
2. Interpretation
for 1 ≤ m ≤ M, at each (x, t). That is, −E{(q + )q} + E{(q + )(q + )T } s = 0, or −E(qq) + {E(qqT ) + E(T )} s = 0
(2.2.19)
since E(q) = 0 by assumption. In detail, (2.2.19) is −Cq (x, t, xn , tn ) +
M Cq (xn , tn , xm , tm ) + Cn,m sm (x, t) = 0
(2.2.20)
m=1
for 0 ≤ x ≤ L, 0 ≤ t ≤ T and 1 ≤ n ≤ M. Solving (2.2.20) for sn yields sn (x, t) =
M
{Cq + C }−1 n,m C q (x, t, x m , tm ),
(2.2.21)
m=1
where the superscript “−1” indicates a matrix inverse, and {Cq }n,m = Cq (xn , tn , xm , tm ). These s(x, t) are the optimal interpolants. They do not depend upon u F or d, but do depend upon the prior covariances Cq and C . In conclusion our best linear unbiased estimate or BLUE is u L (x, t) = u F (x, t) + (d − u F )T {Cq + C }−1 Cq (x, t).
(2.2.22)
Now compare (2.2.22) with (1.3.24), and recall the ﬁrst line in (1.3.32). We have proved: Generalized inversion (the minimization of the integral penalty functional J [u]) is the same as optimal interpolation (the minimization of the local error variance Ee2L (x, t)) when the solution of the forward model u F is the mean ﬁeld, when the data weight matrix w is the inverse of the data error covariance matrix C , and when the reproducing kernel is the covariance Cq . In particular (see (2.2.8), (2.2.9)), rm (x, t) = (x, t, xm , tm ) = Cv (x, t, xm , tm ) = Cq (x, t, xm , tm ), Rnm = (xn , tn , xm , tm ) = Cv (xn , tn , xm , tm ) = Cq (xn , tn , xm , tm ).
Generalized Inversion is Optimal Interpolation
Note 1. Our model is linear, and the data are pointwise. Note 2. OI is widely used in meteorology and oceanography, for the “analysis” or “mapping” of scalar data when the statistical properties of the ﬁelds are plausibly independent of coordinate origins or orientations, both in space and time. That is, when Cq (x, t, y, s) = Cq (x − y, t − s),
(2.2.23)
2.3 The reduced penalty functional
41
for example. Such covariances need only involve a few parameters, which should be reliably estimable from reasonably large sets of data. However, we are increasingly obliged to admit that different ﬁelds are dependent, on dynamical or chemical or biological grounds, so we should use multivariate or vector forms of OI. Moreover, planetaryscale and coastal circulation are obviously statistically inhomogeneous, while the endless emergence of trends suggests statistical nonstationarity. That is, (2.2.23) is false. OI may be generalized to the multivariate, inhomogeneous and nonstationary case provided that there are credible prior estimates for all the parameters in the covariances of the ﬁelds being mapped. We hope that our dynamical models are getting so faithful to the larger scales that model errors like f must be limited to the smaller scales at which (2.2.23) may be plausible. Thus, we should only need to estimate C f (x, t, y, s) = C f (x − y, t − s). We may then use generalized inversion to generate, in effect, the inhomogeneous and nonstationary multivariate equivalents of Cv = Cv (x, t, y, s) and then perform, in effect, an OI of the data. A serious caution must now be offered. It is misleadingly easy to declare that the dynamical error f , initial error i, etc., are random variables belonging to some ensemble, and to manipulate their ensemble moments E f , Ei, E( f f ), E( f i), etc. It is much harder to devise a credible method for estimating these moments. The ﬁelds must clearly be statistically homogeneous at least in one spatial direction or in time, but the presence of spatial or climatological trends makes such homogeneity far from clear. Worse, our dynamical models have already been Reynoldsaveraged or subgridscaleaveraged, so f in particular is already an average of a certain kind. The statistical interpretation of variational assimilation requires, therefore, a second randomization. This difﬁcult issue will be discussed in greater detail in §5.3.7. 2.3
The reduced penalty functional
2.3.1
Inversion as hypothesis testing
Inverse methods enable us to smooth data using a dynamical model as a constraint. Equally, the methods enable us to test the model using the data. The concept of a model is extended here to include not only equations of motion, initial conditions and boundary conditions, but also an hypothesis concerning the errors in each such piece of information. If the model fails the test for a given data set, then the interpolated data or “analysis ﬁeld” is suspect. If the test is failed repeatedly for many data sets, then the hypothesis is suspect. This would be an unsatisfactory state of affairs from the point of view of the ocean analyst or ocean forecaster, but should please the ocean modeler: something new would have been learned about the ocean, namely, that the errors in the dynamics, initial conditions or boundary conditions had been underestimated. Lagrange multipliers make it possible (exercise!) to distinguish between forcing errors and additive components of parameterization errors.
42
2. Interpretation
The hypothesis test is based on the statistical interpretation developed in the previous section. The derivation of the test is very short, but the realization that inverse methods enable model testing is so important that a separate section is warranted.
2.3.2
Explicit expression for the reduced penalty functional
First, recall from (1.3.26) and §2.2.3 that the minimum value of the penalty functional J is ˆ = hT P−1 h Jˆ ≡ J [u] = (d − u F )T (R + C )−1 (d − u F ),
(2.3.1)
where the actual or true “ocean circulation” is u = u F + v,
(2.3.2)
where v is the model response to the random inputs f , i and b, and the data are d = u + ,
(2.3.3)
where is random measurement error. Hence h ≡ d − u F = u + − (u − v) = + v, Eh = E + Ev = 0,
(2.3.4) (2.3.5)
E(hhT ) = E(( + v)( + v)T ) = E(T ) + E(vT ) + E(vT ) + (vvT ) = C + 0 + 0 + R = P + E(vvT ),
(2.3.6)
which statements are parts of, or consequences of, our null hypothesis H0 deﬁned by (2.2.1)–(2.2.3). 1 Now deﬁne P 2 , which is meaningful since P is positivedeﬁnite and symmetric, and hence deﬁne k ≡ P− 2 h. 1
(2.3.7)
Then Ek = P− 2 Eh = 0, 1
E(kk ) = P
− 12
E(hh ) P
=P
− 12
PP− 2
T
T
(2.3.8) − 12
1
= I.
(2.3.9)
That is, E(kn km ) = δnm .
(2.3.10)
2.3 The reduced penalty functional
2.3.3
43
Statistics of the reduced penalty: χ2 testing
The scaled, prior data misﬁts k1 , . . . , k M are zeromean, uncorrelated, unitvariance random variables and 1 1 Jˆ = hT P−1 h = kT P 2 P−1 P 2 k
= kT k = k12 + · · · + k 2M .
(2.3.11)
2 Therefore Jˆ = χ M , the chisquared random variable with M degrees of freedom. Or is it? To be precise, 2 2 χM = x12 + · · · + · · · + x M ,
(2.3.12)
1 p(xm ) = (2π )− 2 exp −xm2 /2 ,
(2.3.13)
where the pdf for xm is
1 ≤ m ≤ M. That is, each xm is a Gaussian random variable having zero mean and unit variance: xm ∼ N (0, 1) (Press et al., 1986). If we had included in H0 the assumption that f , i, b and were Gaussian, then by linearity h and hence k would also be Gaussian. If we do not make that assumption, we may invoke the central limit theorem when M is large, to infer that kn =
M
(P− 2 )nm h m ∼ N (0, 1) 1
(2.3.14)
m=1
as M → ∞. But roughly, if H0 is true then 2 Jˆ = χ M .
So we have a chisquared test for our null hypothesis. Now 2 2 2 2 2 2 E χM = 2M. = M, var χ M ≡ E χM − E χM
(2.3.15)
(2.3.16)
If we perform the inversion a number of times with different data, and ﬁnd that our 2 sample distribution has signiﬁcantly bigger ﬁrst or second moments than those of χ M , then we should reject H0 . We would have learned something about the ocean, from the data. Speciﬁcally, we would have learned that the ocean differs from the model, by more than we had hypothesized. Recall that J is inversely proportional to C f , etc.
Exercise 2.3.1 Show that (i)
J F ≡ J [u F ] = hT C−1 h,
(2.3.17)
44
(ii) (iii)
2. Interpretation
T ˆ Jˆ mod ≡ uˆ − u F , uˆ − u F = βˆ R β, T −1 ˆ Jˆ data ≡ (uˆ − d)T C (uˆ − d) = βˆ C β,
(2.3.18) (2.3.19)
where Pβˆ = h. Note 1. J F is “only data misﬁt”. Note 2. In general, Jˆ mod = Jˆ data .
Exercise 2.3.2 Show that (i) (ii) (iii)
1 − 12 2 EJ F = Tr C− + M, RC 1 −1 1 E Jˆmod = Tr R 2 P R 2 , 1 1 E Jˆdata = Tr C2 P−1 C2 .
(2.3.20) (2.3.21) (2.3.22)
Note 1. Usually EJ F E Jˆ = M. Note 2. In general, E Jˆmod = E Jˆdata . Note 3. In order to devise a rigorous and objective test for an ocean model, we have extended the deﬁnition of a model to include an hypothesis about the statistics of the errors in the dynamics, in the initial conditions and in the boundary conditions as well as in the data.
Exercise 2.3.3 Give meanings to the lefthand sides of (2.3.17)–(2.3.22).
Exercise 2.3.4 (Bennett et al., 2000) Show that (i) (ii) (iii)
2 −1 var (J F ) ∼ 2 Tr (C−1 P C ), var (Jˆ data ) ∼ 2 Tr (C P−2 C ),
var (Jˆ mod ) ∼ 2 Tr [(I − P
as M → ∞.
− 12
C P
(2.3.23) (2.3.24) − 12 2
) ]
(2.3.25)
Remark It is difﬁcult to develop a credible null hypothesis H0 . In particular it is difﬁcult to ˆ the resulting inverse estimate or develop the covariances C f , etc. It follows that u, analysis of the circulation, also lacks credibility. It is a misconception, however, to view inversion as “garbage in, garbage out”. Rather, inversion puts the hypothesis to the test. Forward modeling is no less exposed to the charge of “garbage in, garbage
2.4 General measurement
45
out”: it tests the nullest of null hypotheses, namely, that the dynamical errors, initial errors and boundary errors are all zero, which is the rankest of garbage.
2.4
General measurement
2.4.1
Point measurements
Our data thus far have been direct measurements of the circulation ﬁeld u at isolated points in space and time: dm = u(xm , tm ) + m
(2.4.1)
for 1 ≤ m ≤ M, where dm is the datum and m the measurement error. We shall now consider more general measurements (Bennett, 1985, 1990). 2.4.2
Measurement functionals
First note that a map sending a ﬁeld u into a single real number u(z, w) is an example of a functional u → L[u] = u(z, w).
(2.4.2)
In general u may be a vector ﬁeld of velocity components, pressure, temperature etc., but a measurement of u produces a single number. If the ﬁeld is a streamfunction ψ, and the datum is the meridional component of velocity collected from a current meter, then the appropriate functional is ∂ψ (z, w); (2.4.3) ∂x if the ﬁeld is sealevel elevation h, and the datum is the vertical acceleration of a waverider buoy, then ψ→
∂ 2h (z, w); (2.4.4) ∂t 2 if the ﬁeld is ﬂuid velocity u along a zonal acoustic path, and the datum derives from reciprocalshooting tomography, then h→
z2 u→
u(x, w) d x;
(2.4.5)
z1
if the ﬁeld is sealevel elevation, and the datum is collected by a radar beam, then T h→
L dt
0
d x K (x, t)h(x, t); 0
(2.4.6)
46
2. Interpretation
ﬁnally, if the ﬁeld is stratospheric temperature θ and the datum is a radiative energy ﬂux, then by Stefan’s law, θ → θ 4 (z, w).
(2.4.7)
Note 1. Examples (2.4.2)–(2.4.6) are linear: L[au + bv] = aL[u] + bL[v]
(2.4.8)
for any ﬁelds u, v and real numbers a, b. Note 2. Example (2.4.7) is nonlinear. Note 3. Each of (2.4.2)–(2.4.5) can be expressed as (2.4.6): K (x, t) = δ(x − z)δ(t − w) K (x, t) = −δ (x − z)δ(t − w)
K (x, t) = δ(x − z)δ (t − w)
for
(2.4.2),
for for
(2.4.3), (2.4.4).
Exercise 2.4.1
Find K for (2.4.5).
2.4.3
Representers for linear measurement functionals
The penalty functional may be now expressed as J [u] = u − u F , u − u F + (d − L[u])T C−1 (d − L [u]),
(2.4.9)
where LT = (L1 , . . . , L M ) indicates M linear measurement functionals. Note that in earlier sections, Lm [u] ≡ u(xm , tm ).
(2.4.10)
Furthermore, the Riesz representation theorem establishes that if u → Lm [u] is a bounded (sup Lm [u] /  u  < ∞) linear functional acting on a Hilbert space, then there is an element (a ﬁeld) rm in the space such that u, rm = Lm [u]
(2.4.11)
rm (x, t) = Lm (y,s) [(x, t, y, s)],
(2.4.12)
for any ﬁeld u.
Exercise 2.4.2 Verify that
2.4 General measurement
47
where is the reproducing kernel, and the subscripts (y, s) indicate that Lm acts on as a ﬁeld over (y, s), for each (x, t). Recall that is the representer for evaluation of u at (y, s). In fact, show that rm satisﬁes ∂rm ∂rm +c = C f • αm , ∂t ∂x
(2.4.13)
subject to rm = Ci ◦ αm at t = 0, and rm = cCb ∗ αm at x = 0, where −
∂αm ∂αm −c = Lm (y,s) [δ(x − y)δ(t − s)], ∂t ∂x
subject to αm = 0 at t = T , and αm = 0 at x = L.
(2.4.14)
We may now write J [u] = u − u F , u − u F + (d − u, r)T C−1 (d − u, r)
(2.4.15)
and, as before, the minimizer is uˆ = u F + hT P−1 r,
(2.4.16)
h ≡ d − L[u F ]
(2.4.17)
R = r, rT = L[rT ] = “L L[]L LT ”.
(2.4.18)
where
and P = R + C , where
From now on we shall assume that r, with its adjoint ﬁeld α, represents a general linear measurement functional. We shall reserve the notation and nomenclature of the rk = (x, t, z, w), with its adjoint ﬁeld the Green’s function γ = γ (x, t, z, w), for the evaluation functional (2.4.2).
Exercise 2.4.3 Derive the Euler–Lagrange equations for extrema of (2.4.9). In particular, show that the generalization of (1.3.1) is −
∂λ ∂λ ˆ . −c = LT [δδ]C−1 (d − L [u]) ∂t ∂x
(2.4.19)
48
2. Interpretation
2.5
Array modes
2.5.1
Stable combinations of representers
We have seen that, amongst all free and forced solutions of the forward model, the observing system or “array” only detects the representers. We now ask: are some combinations of representers more stably detected than others?
2.5.2
Spectral decomposition, rotated representers
Assume general linear measurement functionals L = (L1 , . . . , L M )T : u → L[u] ∈ R M . That is, L maps the ﬁeld u linearly into the M real numbers L[u]. The data d are of the form d = L[u] + , where is the vector of measurement errors. The representer matrix is Rnm = Ln (x,t) Lm (y,s) [(x, t, y, s)]
(2.5.1)
for 1 ≤ n, m ≤ M, where Ln (x,t) acts on (x, t, , ), etc. In vector notation, R = LL LT .
(2.5.2)
Recall again that the reproducing kernel is also the covariance Cv : see §2.2, Exercise 2.2.1. The minimization of the penalty functional J , deﬁned by (1.5.7), reduces to the solution of the Mdimensional linear system Pβˆ ≡ (R + C )βˆ = h ≡ d − L[u F ],
(2.5.3)
where u F is the solution of the forward model. The representer matrix R depends upon the dynamics, the prior covariances C f , Ci and Cb for dynamical, initial and boundary residuals, and upon the array L, while C is the covariance of measurement errors. Thus P encapsulates all of our prior knowledge of the ocean in general but does not depend upon the prior estimates of forcing, initial and boundary values F, B and I , provided the dynamics and measurement functionals are linear. The symmetry and positive deﬁniteness of P implies the spectral decomposition P = Z ΦZT ,
(2.5.4)
where Z is orthogonal: ZZT = ZT Z = I, and Φ is diagonal: Φ = diag (φ1 , . . . , φ M ), where φ1 ≥ · · · ≥ φ M > 0. Let L be a rotated vector of measurement functionals: L ≡ ZT L ,
(2.5.5)
and deﬁne the rotated representers r = r (x, t) by r ≡ ZT r = ZT L[].
(2.5.6)
2.5 Array modes
49
These are the array modes (Bennett, 1985, 1992). In particular,
R = L L L T = ZT LL LT Z = ZT RZ,
(2.5.7)
while C = ZT C Z.
(2.5.8)
P = R + C = ZT PZ = Φ,
(2.5.9)
Hence
which is diagonal. The rotated representer coefﬁcients βˆ then obey
P βˆ = h ,
(2.5.10)
h = ZT h.
(2.5.11)
where
2.5.3
Statistical stability, clipping the spectrum
The solution for βˆ is trivial, since P = Φ is diagonal:
h βˆ m = m φm
(2.5.12)
for 1 ≤ m ≤ M. We may deduce from (2.3.5) and (2.3.6) that
Eh = 0, E(h h T ) = P = Φ,
(2.5.13)
so E βˆ m = 0,
E((βˆ m )2 ) =
E((h m )2 ) φm = 2 = φm−1 . 2 φm φm
(2.5.14)
That is, the estimated array mode coefﬁcients βˆ m have greater variance if the corresponding eigenvalue φm is smaller; (2.5.12) shows the inverse to be unstable if the prior data misﬁt h projects signiﬁcantly onto eigenvectors of P having very small eigenvalues. Such projections should be discarded for m > m c , where m c is some cutoff. The exact inverse is uˆ = u F + rT βˆ = u F + rT ZZT βˆ = u F + rT βˆ = u F + rT Φ−1 h ,
(2.5.15)
or ˆ u(x, t) = u F (x, t) +
M m=1
rm (x, t)φm−1 h m ,
(2.5.16)
50
2. Interpretation
and so the stabilized approximation is ˆ u(x, t) ∼ = u F (x, t) +
mc
rm (x, t)φm−1 h m .
(2.5.17)
m=1 Array modes rm c +1 , . . . , r M have been made redundant.
Note 1. The components of the vector of rotated measurement functionals need not correspond to individual elements in the array. They correspond to linear combinations of the elements. Note 2. If we arbitrarily make P more diagonally dominant: P → P + σ 2 I,
(2.5.18)
where σ 2 is additional, independent measurement error variance, then the eigenvalues of P become φ1 + σ 2 , . . . , φ M + σ 2 , which all exceed σ 2 . Thus (2.5.18) would seem to stabilize the inverse. Spectral decomposition (2.5.4), rotation (2.5.6) or clipping (2.5.17) would not be required. However, P and P + σ 2 I have the same eigenvectors, so the array modes are unaffected by (2.5.18). The modes rm c +1 (x, t), . . . , r M (x, t) usually have very ﬁne structure, ˆ and retaining them at almost any level yields a “noisy” inverse u(x, t). It is better to clip the spectrum of P than to make P more diagonally dominant. Note 3. The construction of array modes is essentially an analysis of the condition or stability of the generalized inverse of the model plus array, that is, the stability of the minimization of the penalty functional denoted by (1.5.7), (1.5.9) or (2.1.2). There are two major steps in constructing the inverse. The ﬁrst is the discard (2.1.11) of all the unobservable ﬁelds (2.1.8); it is effected by admitting only solutions of the Euler–Lagrange equations. The second step is the solution of the ﬁnitedimensional linear system denoted as (2.1.15) or (2.5.3). Once this system is solved, the coupling in the Euler–Lagrange equations is resolved and the generalized inverse is ﬁnally obtained by the explicit assembly of (1.3.24), or equivalently by a backward integration followed by a forward integration (see §3.1.2). The dimension of the algebraic system is M, the total number of data. The condition of the system is determined by the M eigenvalues of the coefﬁcient matrix P = R + C . The essential point is that the condition of the inverse is determined without ﬁrst making a numerical approximation to the model using, say, ﬁnite differences; the condition is determined at the continuum level. That is, the condition is set by the partial differential equations, initial conditions, and boundary conditions of the model, by the measurement functionals for the observing system or array, and by the form and weighting of the penalty functional (the actual inputs to the model: internal forcing, initial values, boundary values and data values, have no inﬂuence; the stability of the inverse is its sensitivity to them as a class). The inevitable numerical approximation will indeed modify the null space of unobservable
2.6 Smoothing norms, covariances and convolutions
51
ﬁelds somewhat, and will also alter the eigenvalues, especially the smallest, but these effects are spurious and are suppressed in practice by physical diffusion in the dynamics, by convolution with the covariances in the Euler–Lagrange equations, and by the measurement error variance which has a stabilizing inﬂuence in general. Nevertheless, the continuum and discrete analyses of condition make for an interesting comparison. They may be found in Bennett (1985) and Courtier et al. (1993), respectively. To end with a caution, it is imperative to realize that the array modes and assessment of conditioning depend not only upon the dynamics of the ocean model and the structure of the observing system or array, but also upon the hypothesized or prior covariances of the errors in the model and observing system. If subsequent testing of the hypothesis, using data collected by the array, leads to a rejection of the hypothesis, then the array assessment must also be rejected. Model testing and array assessment are inextricably intertwined. Examples will be presented in Chapter 5. For another approach to array design, see Hackert et al. (1998).
2.6
Smoothing norms, covariances and convolutions
2.6.1
Interpolation theory
The mathematical theory of interpolation is very old. It attracted the attention of the founders of analysis, including Newton, Lagrange and Gauss. The subject was in an advanced state of development by 1940; it then experienced a major reinvigoration with the advent of electronic computers. See Press et al. (1986; Section 2) for a neat outline of common methods, and Daley (1991; Chapter 2) for an authoritative account of methods widely used in meteorology and oceanography. What follows here is a brief outline of the theory attributed to E. Parzen, linking analytical and statistical interpolation. Aside from offering deeper insight into penalty functionals, the theory enables us to design and “tune” roughness penalties essentially equivalent to prescribed covariances (and vice versa). This is of critical importance if one intends, either out of taste or necessity, to minimize a penalty functional by searching in the control subspace rather than in the data subspace. The former search requires roughness penalties or weighting operators; the latter search exploits the Euler–Lagrange equations which incorporate covariances. It has been argued in §2.1 that the datasubspace search is in principle highly efﬁcient, but this efﬁciency will be wasted if the convolutionlike integrals of the covariances and adjoint variables appearing in the Euler–Lagrange equations cannot be computed quickly. Fast convolution methods for standard covariances are given here; the methods are critical to the feasibility of datasubspace searches and hence generalized inversion itself. The section ends with some technical notes on rigorous inferences from penalty functionals, and on compounding covariances.
52
2.6.2
2. Interpretation
Leastsquares smoothing of data; penalties for roughness
Let us set aside dynamics for now. Just consider interpolating some simple data d1 , . . . , d M , which are erroneous measurements of the scalar ﬁeld u = u(x) at the points x1 , . . . , x M . For simplicity, assume that x is planar: x = (x, y). We may deﬁne a quadratic penalty functional by J0 = J0 [u] = W0 u 2 dx + w u − d 2 , (2.6.1) D
where D is some planar domain, W0 and w are positive weights, and u = (u(x1 ), . . . , ˆ u(x M ))T . If uˆ = u(x) is an extremum of J , then the calculus of variations implies that ˆ W0 u(x) = −wδ T (uˆ − d),
(2.6.2)
where δ T = δ T (x) = (δ(x − x1 )δ(y − y1 ), . . . , δ(x − x M )δ(y − y M )). So the “smallest” ﬁeld that “nearly” ﬁts the data is a crop of deltafunctions. This is hardly useful. We would prefer a smoother ﬁeld, so we should penalize the roughness of u, using J1 [u] = W1  ∇u 2 dx + w u − d 2 . (2.6.3) D
Extrema of J1 satisfy W1 ∇ 2 uˆ = wδ T (uˆ − d).
(2.6.4)
So the ﬁeld of least gradient which nearly ﬁts the data is a crop of logarithms: recall that ∇ 2 ln x = −(2π )−1 δ(x)δ(y). What’s more, the solution of (2.6.4) is undeﬁned up to harmonic functions (∇ 2 v = 0) such as bilinear functions, which may or may not be ﬁxed by boundary conditions. Logarithmic singularities are most likely undesirable, so we are led to consider 2 2 2 2 2 & % ∂ 2u ∂ u ∂ u 2 2 J2 [u] = dx, W0 u + W1 ∇u + W2 +2 + 2 ∂x ∂ x∂ y ∂ y2 + w  u − d 2 ,
(2.6.5)
which has extrema satisfying W2 ∇ 4 uˆ − W1 ∇ 2 uˆ + W0 uˆ = −wδ T (uˆ − d).
(2.6.6)
Solutions of (2.6.6) behave like  x − xm 2 ln  x − xm  for x near xm . This is usually acceptable. Evidently, any desired degree of smoothness may be achieved by imposing a sufﬁciently severe penalty for roughness. Note that the homogeneous equation corresponding to (2.6.6), subject to suitable boundary conditions, has only the trivial solution: uˆ ≡ 0.
2.6 Smoothing norms, covariances and convolutions
2.6.3
53
Equivalent covariances
Now consider v, the Fourier transform of u: v(k) = u(x)eik·x dx,
(2.6.7a)
where the range of integration is the entire plane and k = (k, l). The inverse transform is −2 u(x) = (2π ) v(k)e−ik·x dk. (2.6.7b) The penalty functional (2.6.5) is equivalent to −2 J2 [u] = (2π ) (W0 + W1  k 2 + W2  k 4 )  v 2 dk + w  u − d 2 ,
(2.6.8)
provided we assume that the domain D is the entire plane. Let the inverse transform of the reciprocal of the roughness weight in (2.6.8) be −2 C(x) = (2π) (W0 + W1 k2 + W2 k4 )−1 e−ik·x dk. (2.6.9) After some calculus, it may be seen that (2.6.8) becomes J2 [u] = u(x)W (x − x )u(x ) dx dx + w  u − d 2 , where
W (x − x )C(x − x ) dx = δ(x − x ).
(2.6.10)
(2.6.11)
Thus there is close relationship between roughness penalties as in (2.6.5), and “nondiagonal sums” such as in (2.6.10). The latter penalty is in turn related to statistical estimation of a ﬁeld having zero mean, and covariance u(x)u(x ) = C(x − x ).
(2.6.12)
Exercise 2.6.1 Verify all the calculus sketched above, and show that C as deﬁned in (2.6.9) only depends upon x. That is, the random ﬁeld u is isotropic.
Exercise 2.6.2 If D is bounded, what boundary conditions must uˆ satisfy, in order to be an extremum of (2.6.5)?
Exercise 2.6.3 (Wahba and Wendelberger, 1980) Express uˆ in terms of representers. What is the associated inner product?
54
2. Interpretation
Figure 2.6.1 Power spectrum. 1
W0 P(k)
1
(2W0)
k1/2 = l 1
k
How should we choose W0 , W1 , and W2 ? The inverse transform (2.6.9) yields, in particular, the hypothetical variance of u(x): (W0 + W1  k 2 + W2  k 4 )−1 dk, u(x)2 = C(0) = (2π)−2 (2.6.13) hence W0 may be chosen to set the variance once W1 /W0 and W2 /W0 have been chosen. For example, let us assume that W1 /W0 = 0,
W2 /W0 = l 4
(2.6.14)
for some length scale l. Then the hypothetical power spectrum of u(x) is P(k) = W0−1 (1 + k 4l 4 )−1 ,
(2.6.15)
where k = k. Deﬁning the halfpower point k 12 by P(k 12 )/P(0) =
1 2
(2.6.16)
(see Fig. 2.6.1), we ﬁnd that k 12 = l −1 .
(2.6.17)
The functional (2.6.5), with parameters obeying (2.6.14), penalizes scales shorter than l (k k 12 = l −1 ) and ﬁts the data more closely if W0 w.
Exercise 2.6.4 The “bellshaped” covariance C(x) = exp(− x 2l −2 )
(2.6.18)
is commonly used in optimal interpolation. Is there a corresponding smoothing norm, of the kind in (2.6.5)? In summary, there are at least two ways of implementing leastsquares smoothing: with covariances or with smoothing norms. These can be precisely or imprecisely
2.6 Smoothing norms, covariances and convolutions
55
matched, by choice of functional forms and parameters. It will be seen that the choice of implementation can be a matter of major convenience.
2.6.4
Embedding theorems
(The following two sections may be omitted from a ﬁrst reading.) We began §2.6 with a discussion of quadratic penalty functionals used in the smoothing of data. It was seen ˆ that the smoothing ﬁeld u(x) could have unacceptably singular behavior near the data points if the “smoothing norm” in the penalty functional were not chosen appropriately, that is, if the functional did not penalize derivatives of u(x) of sufﬁciently high order. This was demonstrated by examining the solution of the Euler–Lagrange equation for ˆ close to the data points. The examination was feasible since the functionals were u, quadratic in u and hence the Euler–Lagrange equations were linear, but we need not restrict ourselves in principle to quadratic functionals. There are powerful, theoretical guides that relate the mathematical smoothness of the estimate uˆ to the differential order and algebraic power of the “smoothing norm” in the penalty functional. What follows is the crudest sketch of these socalled “embedding theorems” (Adams, 1975). Let us suppose that the function u = u(x) behaves algebraically near the point x0 : u(x) ∼ K r α
(2.6.19)
for small r , where r = x − x0 , K is a positive constant and α is a positive or negative constant. The point x is in ndimensional space: x R n . We unrigorously infer that any mth order partial derivative of u is also algebraic near x0 , with (m) D u(x) ∼ K r α−m , (2.6.20) where K is another positive constant. Hence if we raise D (m) u to the power p and integrate over a bounded domain D that includes x0 , then
...
(m) D u(x) p dx ∼ K
D
R r (α−m) p+n−1 dr,
(2.6.21)
0
where R is the radius of D. The integral on the rhs of (2.6.21) is ﬁnite, provided (α − m) p + n − 1 > −1.
(2.6.22)
That is, if the integral on the lhs of (2.6.21) is ﬁnite, then u(x) ∼ K r α < K r m−n/ p
(2.6.23)
for small r . Provided mp > n > (m − 1) p, a rigorous treatment (Adams, 1975, p. 98) would replace the conclusion (2.6.23) with the more conservative inequality
 u(x) − u(x0 )  < K x − x0 λ ,
(2.6.24)
56
2. Interpretation
where 0 < λ ≤ m − n/ p.
(2.6.25)
If we were to include a term like the lhs of (2.6.21) in our smoothing norm, and were to ﬁnd the uˆ that minimizes the penalty functional, then we could conclude that the lhs of (2.6.21) would be ﬁnite, and hence (2.6.24) must hold. The positivity of λ in (2.6.25) ensures that u is at least continuous at x0 . If λ exceeds unity, then we can be sure that u is differentiable at x0 , and so on: λ > k implies D (k) u is continuous at x0 . For example, suppose n = 2 (we are in the plane: x = (x, y)); suppose m = 2 (we include second derivatives) and suppose p = 2 (we have a quadratic smoothing norm as in (2.6.5)); then mp = 4 > n = 2 > (m − 1) p = 2,
(2.6.26)
and so we cannot even be sure that u is continuous. Nevertheless, we learned from the Euler–Lagrange equation (2.6.6) that u ∼ K r 2 ln r , which is actually differentiable. Thus, the “embedding theorem” estimate of smoothness given in (2.6.25) is very conservative. The theorem would have us choose p = 1.9, mp = 3.8 > n = 2 > (m − 1) p = 1.9.
(2.6.27)
Such a fractional power would make the calculus of variations very awkward, but the penalty functional would be well deﬁned and would have a minimum uˆ with guaranteed continuity. 2.6.5
Combining hypotheses: harmonic means of covariances
We have been considering penalty functionals, schematically of the form J [u] = (Mu) ◦ C −1 f ◦ (Mu) + · · · ,
(2.6.28)
where C f is the hypothesized covariance of Mu, M being some linear differential operator or linear model operator in general. We might also hypothesize that Bu is the covariance of u, in which case we could form the penalty functional −1 J [u] = (Mu) ◦ C −1 f ◦ (Mu) + u ◦ B f ◦ u + · · · .
(2.6.29)
What now is the effectively hypothesized covariance for u? Manipulations like integrations by parts yield −1 (Mu) ◦ C −1 f ◦ (Mu) = u ◦ MC f M ◦ u
=u◦
Cu−1
◦ u,
(2.6.30) (2.6.31)
where Cu = M −1 C f M −1 .
(2.6.32)
2.6 Smoothing norms, covariances and convolutions
57
Think of Cu as the covariance of solutions of the model Mu = f , where f has covariance C f . We can now identify the effectively hypothesized covariance: J [u] = u ◦ Cu−1 ◦ u + u ◦ Bu−1 ◦ u + · · · =u◦
A−1 u
◦ u,
(2.6.33) (2.6.34)
where −1 Au = Cu−1 + Bu−1 is the harmonic mean of the two covariances Cu and Bu .
(2.6.35)
Chapter 3 Implementation
It is a long road from deriving the formulae for the generalized inverse of a model and data to seeing results. First experiments (McIntosh and Bennett, 1984) involved a linear barotropic model separated in time, simple coarselyresolved numerical approximations, a handful of pointwise measurements of sea level and a serial computer. Contemporary models of oceanic and atmospheric circulation involve nonlinear dynamics and parameterizations, advanced highresolution numerical approximations, vast quantities of data often of a complex nature, and parallel computers. Chapter 3 introduces some general principles for travelling this long road of implementation. The ﬁrst principle is accelerating the representer algorithm by task decomposition, that is, by simultaneous computation of representers on parallel processors. The objective may be either the full representer matrix as required by the direct algorithm, or a partial matrix for preconditioning the indirect algorithm. The calculation of an individual representer, or indeed any backward or forward integration, may itself be accelerated by domain decomposition, but this is a common challenge in modern numerical computation (Chandra et al., 2001; Pacheco, 1996) and will not be addressed here. Even without considering the coarse grain of task decomposition or the ﬁne grain of domain decomposition, the direct and indirect representer algorithms for linear inverses are highly intricate. Schematics are provided here in the form of “time charts”. Dynamical errors and input errors may be correlated in space or in time or in both. Error covariances must be convolved with adjoint variables. This is a massive task if four dimensions are involved and the numerical resolution is ﬁne. Fast convolutions are critical to the scientiﬁc purpose of leastsquares inversion, which is the testing of hypotheses about model errors. Posterior error statistics are equally essential, and are also massively expensive to compute and store in full detail. These statistics need not
58
3.1 Accelerating the representer calculation
59
be computed with the same precision as the inverse itself, as they are only used for rough assessment of the likely accuracy of the inverse. Storageefﬁcient Monte Carlo algorithms permit computations of selected statistics with adequate reliability, on the same grid as the forward model if so desired. Nonlinearity can only be overcome by iteration, but there is no unique way to iterate. This is a blessing in disguise, as certain choices for functional iterations can lead to linear, unbounded instability. No functional linearization yields statistical linearization, so signiﬁcance tests and posterior error covariances that assume statistical linearity must be used with caution. Finally, crude parameterizations of unresolved natural processes may not be functionally smooth, thereby precluding variational assimilation. This obstacle should in principle be overcome by ﬁddling with the unnatural parameterization. Experience with trivial models suggests that we have much to learn.
3.1
Accelerating the representer calculation
3.1.1
So many representers . . .
The representer algorithm provides an explicit solution of linear Euler–Lagrange equations, and hence leastsquares generalized inverses of overdetermined linear forward problems. There is one representer for each excess datum, and two model integrations are required (one backward, one forward) in order to construct each representer. (Note that we may regard the initial values and boundary values for the forward problem as data having exactly the same status as the ﬁnite set of measurements that overdetermine the forward problem; indeed, we may in principle envisage measurements obtained continuously along a track, and we shall in Chapter 6 consider specifying boundary values of too many components of a vector ﬁeld.) It is impractical to compute every representer if their number is very large. There are rational approaches to reducing their number, as will be indicated in Chapter 5, but such approximations may not be necessary. It is possible to compute the representer solution for the inverse without reducing the number of representers, and without signiﬁcant numerical approximation beyond that already implied by the numerical model. This technical advance has allowed the inversion of large data sets, with complex models imposed as weak constraints. 3.1.2
Openloop maneuvering: a time chart
Recall again from §1.3.3 the representer solution for the inverse: ˆ u(x, t) = u F (x, t) +
M
βˆ m rm (x, t),
(3.1.1)
m=1
where (R + C )βˆ = h ≡ d − L[u F ].
(3.1.2)
60
3. Implementation
Thus our tasks are: (1) integrate the forward model for u F (2) integrate the backward model for α (3) integrate the forward model for r for a total of
... ... ... ...
one integration; M integrations; M integrations; I = 2M + 1 integrations.
The backward and forward parts of the Euler–Lagrange equations are coupled by M numbers uˆ 1 , . . . , uˆ M . The vector coefﬁcient of the impulses in the adjoint equation (2.4.19), or coupling vector, is actually ˆ ˆ = C−1 {d − L[u F ] − RT β} C−1 (d − L[u]) (3.1.3)
T −1 = C−1 {h − R (R + C ) h}
...
= (R + C )−1 h ˆ = β.
(3.1.4)
That is, the coupling vector is the vector of representer coefﬁcients. So we need not store the representer vector ﬁeld r(x, t). We must compute r (x, t), measure it to obtain ˆ integrate the adjoint or backward the representer matrix R = L[rT ], solve (3.1.2) for β, ˆ EL equation (2.4.19) for λ(x, t) and then integrate the forward equation (1.5.14) for ˆ u(x, t). Now the integration count is I = 2M + 3. See Fig. 3.1.1 for a “time chart” implementing this socalled “open loop” version of the representer algorithm. t=0 (1.2.2) & (1.2.4) (1.2.3)
(1.3.12)
(1.3.5)
t=T
uF
(1.3.8) & (1.3.10)
α m,1≤m≤M
(1.3.11) & (1.3.13)
r m,1≤m≤M
(3.1.2)
R, β
(1.3.1) & (1.3.3)
λ
(1.3.4) & (1.3.6)
u^
(1.3.9)
^
(1.3.2)
Figure 3.1.1 Time chart for implementing the representer algorithm with direct calculation of the representer ˆ that is, by explicit or direct construction of the coefﬁcient β, representer matrix R. The heavy vertical arrow on the right indicates the order of execution, which starts at the top. Note that (3.1.1) need not be summed explicitly; once βˆ is known, (3.1.4) resolves the coupling in (1.3.1)–(1.3.6). So in this “open loop” version, λ and hence uˆ may be calculated with one backward integration and one forward integration. The representers rm , 1 ≤ m ≤ M, need not be stored. If the −1 and wb−1 are nondiagonal, then inverse weights W −1 f , Wi (1.3.11)–(1.3.13) and (1.3.4)–(1.3.6) require convolutions as in (1.5.14)–(1.5.16). See also (4.2.1)–(4.2.6) for a statement of the Euler–Lagrange equations for nondiagonal weighting.
3.1 Accelerating the representer calculation
61
Table 3.1.1 Processor work sheet. 1 uF α1 r1
···
m
···
M
uF αm rm
uF αM rM
L[r1 ] L[rm ] L[r M ] (all processors get the other M−1 columns of R) (all processors solve for βˆ = (R + C )−1 (d − L[u F ])) λ λ λ uˆ uˆ uˆ I =5
3.1.3
I =5
I =5
Task decomposition in parallel
The above task is ideally suited to parallel processing (Bennett and Baugh, 1992). Suppose we have M processors 1 , . . . , M . The work sheet is as follows (see Table 3.1.1). The m th processor m calculates the ﬁelds u F , αm and rm , takes all M measurements L1 , . . . , L M of rm to obtain the m th column of R, broadcasts this column to all of the other M − 1 processors, receives the other M − 1 columns in return, assembles R, ˆ then solves the Euler–Lagrange equations by calculating the solves for the vector β, ˆ ﬁeld λ and the ﬁeld u. So I is reduced from 2M + 3 to 5 with an Mprocessor system. There is minimal exchange of data: each processor broadcasts one column of the representer matrix, and receives M − 1 columns in return. Each processor m must have sufﬁcient memory and speed for the computation and storage of u(x, t), for 0 ≤ x ≤ L and 0 ≤ t ≤ T . Note that the calculations of u F , λ and uˆ are Mfold redundant. This permits the programmer to release M − 1 processors during these steps; more importantly it permits a reduction of the number of broadcast messages by a factor of M − 1.
3.1.4
Indirect representer algorithm; an iterative time chart
Generalized inversion reduces exactly to solving the ﬁnitedimensional system (R + C )βˆ = (d − L[u F ]),
(3.1.5)
Pβˆ = h.
(3.1.6)
or simply
A direct solution requires that P and hence R be explicitly known. However, the solution may be obtained iteratively, provided Pψ can be evaluated for any vector ψ. Then a standard iterative solver can convert a ﬁrstguess βˆ 0 into a solution βˆ = P−1 h.
62
3. Implementation
Let us now examine how we could compute Pψ, given any vector ψ. We have Pψ = Rψ + C ψ.
(3.1.7)
The data error covariance matrix C is explicitly known, so the nontrivial problem is the evaluation of Rψ. The following procedure (Egbert et al., 1994; Amodei, 1995; Courtier, 1997) does that without calculating the representers. First, solve the backward model, with coupling vector ψ: −
∂φ ∂φ −c = ψ T L[δδ], ∂t ∂x
(3.1.8)
subject to φ=0
(3.1.9)
φ=0
(3.1.10)
at t = T , and
at x = L. Second, solve the forward model, with adjoint ﬁeld φ(x, t): ∂θ ∂θ +c = C f • φ, ∂t ∂x
(3.1.11)
θ = Ci ◦ φ
(3.1.12)
θ = cCb ∗ φ
(3.1.13)
subject to
at t = 0, and
at x = 0. Comparison of (3.1.8)–(3.1.13), with the equations (2.4.14), (2.4.13) for α and r respectively, shows that φ(x, t) = ψ T α(x, t),
θ (x, t) = ψ T r(x, t) = r(x, t)T ψ.
(3.1.14)
Hence L[θ] = L[rT ]ψ = Rψ,
(3.1.15)
which is just what is needed, at a cost of two integrations; see Fig. 3.1.2.
3.1.5
Preconditioners
If P is the unit matrix, then iterative solution of (3.1.6) should converge to h in one step. Hence iteration on (3.1.6) should in general be accelerated by premultiplying both sides of (3.1.6) with the inverse of a symmetric, positivedeﬁnite approximation to P. That is, solve −1 ˆ P−1 A Pβ = PA h,
(3.1.16)
3.1 Accelerating the representer calculation
t=0 (1.2.3)
t=T
(1.2.2) & (1.2.4) (3.1.8) & (3.1.10)
(3.1.12)
63
(3.1.9)
(3.1.11) & (3.1.13) (3.1.15) & (3.1.7) (1.3.1) & (1.3.3)
(1.3.5)
(1.3.2)
(1.3.4) & (1.3.6)
Figure 3.1.2 Time chart for implementing the indirect representer algorithm. The representer coefﬁcients βˆ are approximated by iterative solution of (3.1.2). Given a previous ˆ the “inner iteration” calculates approximation ψ for β, Pψ = Rψ + C ψ with one backward integration and one forward integration (the representer matrix R is not explicitly constructed). This information is then used to ﬁnd a better ˆ Once βˆ has been approximated with approximation ψ for β. sufﬁcient accuracy, λ and hence uˆ are calculated as in the direct “open loop” algorithm. That is, the sum (3.1.1) is evaluated implicitly by one backward integration and one forward integration.
where PA ∼ = P. Then (3.1.16) may be solved iteratively since, if we can evaluate Pψ for any ψ, then we can also evaluate P−1 A Pψ. There are various choices for the preconditioner PA . (i) (Bennett et al., 1996) We could calculate all the representers quickly and cheaply on a coarse grid. Note that we would still be in effect solving (3.1.6) for the coefﬁcients of the representers on the ﬁne grid, so there would be no loss of ˆ resolution in u(x, t). However, the grid vertices may not coincide very closely with observing sites; the measurement functionals must involve interpolation formulae and these degrade appreciably as the grid gets coarser. That is, PA may be a poor approximation to P and so convergence may not be greatly accelerated. (ii) (Egbert and Bennett, 1996; Egbert, 1997) We could calculate some of the representers on the ﬁne grid. Let RC be the M × K matrix consisting of the ﬁrst K columns of R. That is, R = (RC , RNC ), where the noncalculated matrix RNC is of dimension M × (M − K ). Now RC may be partitioned into an upper K × K block denoted by R11 , and a lower (M − K ) × K block R21 , etc. That is, R=
R11 R21
R12 R22
= (RC , RNC ).
(3.1.17)
64
3. Implementation
The representer matrix is symmetric, therefore R12 is actually known at this T point: R12 = RT21 . An estimate for R22 is R21 R−1 11 R21 , thus R11 RT21 . (3.1.18) RA = T R21 R21 R−1 11 R21 Then PA = RA + C . Note that the ranks of RA and PA are K and M respectively. The effectiveness of this preconditioner depends upon a judicious choice for the K calculated representers, and upon the independence of the measurement errors. (iii) Recall from (2.2.9) and (2.4.18) that the representer matrix is a covariance: R = LCv LT
(3.1.19)
= E{(L[u] − L[u F ])(L[u] − L[u F ]) }. T
(3.1.20)
Thus we may estimate R by Monte Carlo methods. That is, we make pseudorandom samples of L[u − u F ] and then evaluate sample covariances. The issue is: how many samples sufﬁce? Further details on implementation, including a ﬂowchart, may be found in Chapter 5. The issue of sample size will be illustrated in §5.5. 3.1.6
Fast convolutions
We have seen the need to assume “nondiagonal” covariances for dynamical errors; that is, C f (x, t, x , t ) = δ(x − x )δ(t − t ).
(3.1.21)
The covariance appears in the “forward” equation for the inverse estimate, for example (1.5.14): ∂ uˆ ∂ uˆ (x, t) + c (x, t) = F(x, t) + (C f • λ)(x, t), (3.1.22) ∂t ∂x where T L (C f • λ)(x, t) = ds dy C f (x, t, y, s)λ(y, s). (3.1.23) 0
0
Direct evaluation of this integral for each (x, t) would be prohibitively expensive in only one space dimension and time, and even more so in several space dimensions and time. Thus, a crucial requirement for smooth and hence physically acceptable inversions is an efﬁcient algorithm for the evaluation of integrals such as (3.1.23). We shall refer to these loosely as “convolutions”. The following shortcut is very efﬁcient (Derber and Rosati, 1989; Egbert et al., 1994). Assume that the covariance is purely spatial, and is “bellshaped”: C(x, x ) = C0 exp(−x − x 2 /L 2 ),
(3.1.24)
3.1 Accelerating the representer calculation
65
where C0 is a constant. Assume that we wish to evaluate ∞ ∞ b(x) =
C(x, x )a(x ) dx .
(3.1.25)
−∞ −∞
Solve the following pseudoheat equation for θ = θ (x, s): ∂θ = ∇ 2 θ, ∂s
(3.1.26)
θ (x, 0) = a(x).
(3.1.27)
by timestepping, subject to
In two space dimensions, the solution is −1
∞ ∞
θ (x, s) = (4πs)
exp(−x − x 2 /(4s))a(x ) dx .
(3.1.28)
−∞ −∞
So let s = L 2 /4,
(3.1.29)
b(x) = π L 2 C0 θ (x, L 2/4).
(3.1.30)
then
Exercise 3.1.1 Compare the operation counts for numerical integration of (3.1.26), and numerical evaluation of (3.1.25), for one, two and three space dimensions.
Exercise 3.1.2 How might you proceed when the spatial domain is ﬁnite?
If the covariance is inhomogeneous, for example C(x, x ) = V (x) 2 V (x ) 2 exp(−x − x 2 /L 2 ), 1
1
(3.1.31)
where V (x) = C(x, x) is the variance, then proceed as above except that the initial condition becomes 1
θ(x, 0) = V (x) 2 a(x),
(3.1.32)
and the required result is 1
b(x) = V (x) 2 π L 2 θ(x, L 2 /4). Now consider temporal convolution, involving the simple form C(t, t ) = exp(−t − t /τ ).
(3.1.33)
66
3. Implementation
That is, we wish to evaluate T b(t) =
C(t, t )a(t ) dt .
(3.1.34)
0
This is the solution of btt − τ −2 b = −2τ −1 a
(3.1.35)
bt − τ −1 b = 0
(3.1.36)
bt + τ −1 b = 0
(3.1.37)
for 0 ≤ t ≤ T , subject to
at t = 0, and
at t = T . The two point boundaryvalue problem (3.1.35)–(3.1.37) is easily solved as two initialvalue problems. First, solve h t + τ −1 h = −2τ −1 a
(3.1.38)
h=0
(3.1.39)
bt − τ −1 b = h
(3.1.40)
b = −(τ/2)h
(3.1.41)
for 0 ≤ t ≤ T , subject to
at t = 0. Then solve
for 0 ≤ t ≤ T , subject to
at t = T .
Exercise 3.1.3 Show that the order of the two integrations in Exercise 3.1.2 may be reversed, with a modiﬁcation to the terminal conditions.
3.2
Posterior errors
3.2.1
Strategy
ˆ If u is the true circulation, and if we adopt the How good is the generalized inverse u? statistical interpretation of the inverse, then the error u − uˆ has zero mean. There is a closed expression for the covariance of this error, or posterior error covariance. The
3.2 Posterior errors
67
expression involves the covariance of u − u F prescribed a priori in H0 , and all of the representers. An efﬁcient strategy for evaluating this formidable expression is essential. The direct, serial representer algorithm requires the computation of M representers, one per datum. Each computation requires one backward and one forward integration; these may be executed in parallel if resources permit (see §3.1). It has been shown in §2.2 and §2.4 that the mth representer rm (x, t) is in fact the covariance of the mth measurement Lm [v] and the ﬁeld v(x, t) itself, where v is the response of the model to random forcing consistent with the hypothesis H0 . The M representers having been computed, and stored, they may be used to construct error covariances for the inverse estimates fˆ, ıˆ, bˆ ˆ of the forcing, initial values, boundary values and measurements respectively and L[u] (Bennett, 1992, §5.6). Computation of uˆ using representers indirectly, as in §3.1.4, does not yield these posterior error covariances. The indirect approach typically requires about 10% of the effort of the direct approach; such efﬁciency is sometimes achieved by preliminary computation of the representers, either in part on the actual model grid or in the total on a coarser grid. This incomplete covariance information may sufﬁce ˆ as an indication of the reliability of u. Regardless of the implementation of the representer algorithm, that is, either direct ˆ it is possible to make “Monte or indirect solution of the Euler–Lagrange equations for u, Carlo” estimates of just as much covariance information as is required. The level of ˆ but it is satisfactory as an indicator accuracy may be below that used to compute u, ˆ The version of the Monte Carlo algorithm given in §3.2.5 is of the reliability of u. complicated, but it is highly memoryefﬁcient. 3.2.2
Restatement of the “toy” inverse problem
For convenience, let us restate the “toy” problem here. The true ocean circulation u satisﬁes ∂u ∂u (x, t) + c (x, t) = F(x, t) + f (x, t), ∂t ∂x u(x, 0) = I (x) + i(x), u(0, t) = B(t) + b(t),
(3.2.1) (3.2.2) (3.2.3)
where F, I and B are respectively the prior estimates of the forcing, initial values and boundary values (prior to assimilating data), while f , i and b are respectively the unknown errors in those priors. The prior estimate of u is u F , which satisﬁes ∂u F ∂u F (x, t) + c (x, t) = F(x, t), ∂t ∂x u F (x, 0) = I (x, t), u F (0, t) = B(t).
(3.2.4) (3.2.5) (3.2.6)
The data comprise an Mdimensional vector d: d = L[u] + ,
(3.2.7)
3. Implementation
68
where L is a vector of linear measurement functionals and is the vector of the measurement errors. In order to improve upon u F , we make an hypothesis H0 about the unknown errors f , i, b and : E f (x, t) = Ei(x) = Eb(t) = 0,
E = 0; E( f (x, t) f (x , t )) = C f (x, t, x , t ), E(i(x)i(x )) = Ci (x, x ),
E(b(t)b(t ))
= Cb (t, t ),
E(T )
= C ,
E( f b ) = E( f i ) = E(ib ) = 0,
E( f ) = E(i) = E(b) = 0.
(3.2.8)
(3.2.9)
(3.2.10)
That is, we assume that the errors f , i, b and have vanishing means (F, I , B and d are unbiased) and have speciﬁed covariances C f , Ci , Cb and C . Then the posterior estimate uˆ minimizes the estimator −1 −1 T −1 J [u] ≡ f • C −1 f • f + i ◦ C i ◦ i + b ∗ C b ∗ b + C ,
(3.2.11)
where f , i, b and are related to u via (3.2.1)–(3.2.3) and (3.2.7). The symbols •, ◦ and ∗ are deﬁned by !L !T f • g ≡ d x dt f (x, t)g(x, t), 0 0 L ! (3.2.12) i ◦ j ≡ d x i(x) j(x), 0 !T a ∗ b ≡ dt a(t)b(t). 0
−1 The inverse covariances C −1 and Cb−1 are deﬁned in terms of (3.2.12): f , Ci !L !T −1 d x dt C f (x, t, x , t )C f (x , t , x , t ) = δ(x − x )δ(t − t ), 0 0 L ! −1 (3.2.13) d x Ci (x, x )Ci (x , x ) = δ(x − x ), 0 !T −1 dt Cb (t, t )Cb (t , t ) = δ(t − t ). 0
The inverse of C is the standard matrix inverse C−1 : C−1 C = I,
(3.2.14)
where I is the M × M unit matrix. The minimizer of J and optimal estimate of u is ˆ u(x, t) = u F (x, t) + βˆ T r(x, t),
(3.2.15)
3.2 Posterior errors
69
where the representer ﬁelds r = r(x, t) and adjoint variables α = α(x, t) satisfy ∂α ∂α −c = L[δδ], ∂t ∂x α = 0 at t = T,
−
α=0
at
x = L,
(3.2.16) (3.2.17) (3.2.18)
∂r ∂r +c = C f • α, ∂t ∂x r = Ci ◦ α at t = 0,
(3.2.20)
r = cCb ∗ α at
(3.2.21)
x = 0.
(3.2.19)
In (3.2.16), δδ = δ(x − y)δ(t − s) and L acts upon the (y, s) dependence. The representer coefﬁcients βˆ satisfy the linear system Pβˆ ≡ (R + C )βˆ = d − L[u F ] ≡ h,
(3.2.22)
where R = L[rT ] = L[]LT , = (x, t, y, s) being the reproducing kernel (‘rk’), or representer for a point measurement at (y, s).
3.2.3
Representers and posterior covariances
ˆ The error in the state estimate uˆ is deﬁned to be u(x, t) − u(x, t), and we would like to know its mean and covariance. First, let us note that the error in the prior estimate of the state has zero mean: E(u(x, t) − u F (x, t)) = 0,
(3.2.23)
Cu (x, t, x , t ) ≡ E((u(x, t) − u F (x, t))(u(x , t ) − u F (x , t ))),
(3.2.24)
and its covariance is
which is also, as was established in §2.2.3, the rk (x, t, x , t ). The latter is, again, the representer for point measurement at (x , t ). Indeed (Exercise 2.2.1), = γ • C f • γ + γ ◦ Ci ◦ γ + c2 γ ∗ C b ∗ γ ,
(3.2.25)
where γ (x, t, x , t ) is the inﬂuence function or Green’s function for the “toy” model. (The symbol Cv in §2.2.3 has the same deﬁnition as Cu here in §3.2.3; it seems helpful to use different symbols in the two sections.) Notice that , and hence Cu , is determined by the model, and by the prior covariances C f , Ci and Cb . These are, again, the covariances of the errors f , i and b in the prior estimates F, I and B of the forcing, initial and
70
3. Implementation
boundary values, respectively. Recall also that P ≡ R + C = E(hhT ),
(3.2.26)
where h is the prior data misﬁt: h = d − L[u F ].
(3.2.27)
Our main result is a tedious consequence of the above.
Exercise 3.2.1 (Bennett, 1992, §5.6; Xu and Daley, 2000) Show that the state estimate is unbiased: ˆ E(u(x, t) − u(x, t)) = 0,
(3.2.28)
and has as its covariance ˆ ˆ , t ))) Cuˆ (x, t, x , t ) ≡ E((u(x, t) − u(x, t))(u(x , t ) − u(x = Cu (x, t, x , t ) − rT (x, t)P−1 r(x , t ).
(3.2.29)
Recall that the optimal estimates of f , i, b and are fˆ(x, t) = (C f • λ)(x, t),
(3.2.30)
ıˆ(x) = (Ci ◦ λ)(x, 0),
(3.2.31)
ˆ = c(Cb ∗ λ)(0, t) b(t)
(3.2.32)
ˆ ˆ = d − L[u],
(3.2.33)
and
ˆ where λ(x, t) is the weighted residual, or variable adjoint to u(x, t): ∂ uˆ ∂ uˆ λ = C −1 +c −F . f • ∂t ∂x
(3.2.34)
It is readily shown that these all have zero mean: E fˆ = E ıˆ = E bˆ = 0,
E ˆ = 0.
(3.2.35)
The posterior error covariances for f , i, b and follow easily from (3.2.29): C fˆ (x, t, x , t ) ≡ E(( f (x, t) − fˆ(x, t))( f (x , t ) − fˆ(x , t ))) = C f (x, t, x , t ) − sT (x, t)P−1 s(x , t ),
(3.2.36)
where s is the representer residual vector: s≡
∂r ∂r +c ; ∂t ∂x
(3.2.37)
3.2 Posterior errors
71
Cıˆ (x, x ) ≡ E((i(x) − ıˆ(x))(i(x ) − ıˆ(x ))) = Ci (x, x ) − rT (x, 0)P−1 r(x , 0);
(3.2.38)
ˆ ˆ ))) Cbˆ (t, t ) ≡ E((b(t) − b(t))(b(t ) − b(t = Cb (t, t ) − rT (0, t)P−1 r(0, t ),
(3.2.39)
and Cˆ ≡ E(( − )( ˆ − ) ˆ T) = R − R P−1 R −1
= C − C P C .
(3.2.40) (3.2.41)
Examination of (3.2.29)–(3.2.41) shows that, since C f , Ci , Cb and C are prescribed, calculating the M representers r(x, t) yields C fˆ , Cıˆ , Cbˆ and Cˆ . Only Cuˆ is not so available, since that requires Cu = . We do know Cuˆ at data sites: see (3.2.40). Thus the M ˆ fˆ, ıˆ, bˆ and ˆ , and all the posterior error covariances representers give us the estimates u, except Cuˆ . The difference Cu − Cuˆ , or the “explained” covariance, may be expressed in terms of r. In principle, we could calculate Cu (x, t, x , t ) as the rk (x, t, x , t ), that is, by calculating the representer for every point (x , t ), but that is impractical. If the data were sufﬁciently dense, we could interpolate C to ﬁnd Cuˆ between data sites, but that is useful only if L involves just point measurement, and involves point measurement of every component when the state is multivariate (u, v, w, p, etc.).
3.2.4
Sample estimation
We may approximate the prior error covariance Cu using sample averages. Let the prior error be denoted by v(x, t) ≡ u(x, t) − u F (x, t).
(3.2.42)
∂v ∂v +c = f, ∂t ∂x
(3.2.43)
Then
v=i
at t = 0,
(3.2.44)
v=b
at
x = 0.
(3.2.45)
Use pseudorandom number generators to create pseudorandom ﬁelds f (x, t), i(x) and b(t) consistent with the null hypothesis H0 . For example, construct a “whitenoise” ﬁeld w(x) satisfying Ew(x) = 0,
E(w(x)w(x )) = δ(x − x ),
(3.2.46)
72
3. Implementation
then “color” w to obtain a realization or sample for the initial error ﬁeld i: 1 i(x) = Ci2 ◦ w (x) =
L
1
Ci2 (x, x )w(x ) d x ,
(3.2.47)
0
where 1
1
Ci2 ◦ Ci2 = Ci .
(3.2.48)
Then Ei(x) = 0,
E(i(x)i(x )) = Ci (x, x )
(3.2.49)
as in H0 . Samples of f (x, t) and b(t) may similarly be constructed. In computational practice, the real variable x is replaced with a grid xn = nx for some uniform step x. Fast subroutines generate random numbers r independently √ and uniformly distributed in the interval 0 < r < 1. Let s = 2 3(r − 12 ); then Es = 0 √ and E(s 2 ) = 1. Let sn be such a number; and let wn = sn / x. Hence Ewn = 0, and 1 E(wn wm ) = δnm /x ∼ = δ(xn − xm ). Then take i n = m Ci2nm wm . Generate K such samples of i n : i n1 , i n2 , . . . , i nK , and similarly generate f nlk , blk , where l is a time index. Approximate (3.2.43)–(3.2.45) on the (n, l) ﬁnitedifference grid. Integrate numerik cally to obtain samples vnl for k = 1, . . . , K . Then the sample prior error covariance is Cu (xn , tl , x p , tq ) ∼ = K −1
K
k k vnl v pq .
(3.2.50)
k=1 k It is advisable to remove ﬁrst any spurious sample mean of vnl . Armed with this approximation to Cu = , we may evaluate the representers r = L[] and hence all the posterior error covariances (3.2.29), (3.2.36)–(3.2.41).
Note 1. The prior data error covariance C inﬂuences the posteriors Cuˆ , . . . , Cˆ , but the actual data d do not. Note 2. The posteriors Cuˆ , Cıˆ , Cbˆ , and Cˆ given in (3.2.29), (3.2.38), (3.2.39) and (3.2.41) need not be related to a model; Cuˆ is the posterior for the best linear unbiased estimate of u based on a prior u F and data d having errors with zero means and covariances Cu and C , respectively. Recall that r = L[Cu ]. The measurement functionals L must be linear. The posterior (3.2.36) is valid only if f is related to u via a linear model. Note 3. The posterior state estimate uˆ may be expressed in terms of representers calculated as r = L[Cu ], where Cu is a sample covariance. However, a great many samples are needed for this approach to agree accurately with solutions of the representer equations (3.2.16)–(3.2.21) (Bennett et al., 1998). The latter approach also requires many integrations, but the number of sample integrations should actually be compared to the cost of computing a preconditioner for an indirect representer solution as in §3.1.5.
3.2 Posterior errors
73
Storage becomes a serious problem if Cu must be retained in full. It may sufﬁce, for the purposes of indicating error levels, to compute Cu on a much coarser space–time ˆ grid than that used to calculate the state estimate u.
3.2.5
Memoryefﬁcient sampling algorithm
The following algorithm for sample estimates of Cuˆ is memoryefﬁcient but complicated: (i) generate samples of f k , i k and bk , k = 1, . . . , K ; (ii) integrate to ﬁnd samples for v k , k = 1, . . . , K ; (iii) make a sample estimate of the representer matrix: R∼ = R K = K −1
K
L[v k ]L[v k ]T ,
(3.2.51)
k=1
(iv) (v) (vi) (vii)
(note that it is not necessary to store all the K samples in order to evaluate (3.2.51), and note also that the rank of R K is K ); regenerate samples f k , i k and bk , k = 1, . . . , K , identical to those in (i); recompute v k , k = 1, . . . , K ; generate samples of measurement error k , k = 1, . . . , K ; derive sample data misﬁts: hk = L[v k ] + k
(3.2.52)
for k = 1, . . . , K ; (viii) solve for the sample representer coefﬁcients βˆ k : (R + C )βˆ k = hk
(3.2.53)
for k = 1, . . . , K (use the indirect method of §3.1.4 and precondition with (R K + C )); (ix) solve the Euler–Lagrange equations for the sample posterior state estimate uˆ k = u F + (βˆ k )T r
(3.2.54)
for k = 1, . . . , K ; (x) evaluate the sample mean posterior error covariance: Cuˆ (x, t, x , t ) ∼ = K −1
K
(u k (x, t) − uˆ k (x, t))(u k (x , t ) − uˆ k (x , t ))
k=1
(3.2.55) (note that u k = u F + v k ).
74
3. Implementation
This twostage statistical simulation is intricate and requires many more model integrations than does singlestage simulation (3.2.50, etc.), but uses minimal memory without resorting to coarser grids. Note 1. It is not necessary to evaluate (3.2.55) for all (x, t, x , t ); it may be evaluated ˆ as much as is needed, in order to indicate the reliability of u. Note 2. The posterior error means and covariances derive from the prior error moments assumed in H0 (see (3.2.8)–(3.2.10)). Rejection of H0 implies rejection of the posterior moments.
Exercise 3.2.2 Compare the computational requirements and storage requirements of the Monte Carlo algorithms given in §3.2.4 and in §3.2.5.
3.3
Nonlinear and nonsmooth estimation
3.3.1
Double, double, toil and trouble
Linear leastsquares estimation problems may be solved efﬁciently by exploiting their linearity. The null subspace may be suppressed. Its complement, the data subspace, may be spanned with a ﬁnite basis – the representers. Their coefﬁcients may be sought iteratively, and their sum may be formed without constructing each representer. Alas, nonlinearity is intrinsic to geophysical ﬂuid dynamics, while many parameterizations involve functional nonsmoothness that precludes variational analysis. The general approach to smooth nonlinearity is to iterate yet again, leading to sequences of linear leastsquares problems with solutions converging to that of the nonlinear leastsquares problem. Several such “outer” iteration schemes are described here. The most orderly of them – the tangent linearization scheme – has the potential for grossly unphysical behavior. All linearization schemes are potentially unstable, drawing energy from the reference ﬁeld and lacking amplitude modulation of that unstable growth. Practical experience of iterating on nonlinear Euler–Lagrange equations is not so bad: see Chapter 5. It seems that more data lead to faster convergence, while moderate smoothing of sources of linear instability in the adjoint equations can ensure stability at the price of slight suboptimality. Nonlinearity in the dynamics vitiates the statistical analyses of §2.2 and §3.2. Dynamical linearization does not lead to statistical linearization, so the signiﬁcance tests and recipes for posterior error covariances are suspect. In particular, bias can emerge in the inverse estimate of circulation even when none is present in the prior estimate of forcing. Finally, nonsmoothness of the dynamics or the penalty functional precludes variational analysis. However, nonsmoothness is unnatural; it is an admission of poor
3.3 Nonlinear and nonsmooth estimation
75
resolution. Any mathematical “fudge” that removes it is entirely justiﬁed, since the nonsmoothness is itself a fudge. What is really needed is not more theory, but more experience with realistic models and copious data. Nevertheless, here are some introductory analyses.
3.3.2
Nonlinear, smooth dynamics; leastsquares
Consider a nonlinear wave equation: ∂u ∂ + {U (u)u} = F + f, ∂t ∂x
(3.3.1)
subject to an initial condition u(x, 0) = I (x) + i(x)
(3.3.2)
at t = 0, and subject to the boundary condition u(0, t) = B(t) + b(t)
(3.3.3)
at x = 0. As usual, F, I and B are priors, while f , i and b are unknown errors in the priors. The phase speed U is now a known function of the “ocean circulation” u. The form (3.3.1) is not the most general nonlinear wave equation, but it represents nondivergent advection in ocean models. A simple penalty functional is L J [u] = W f
T dt f + Wi
dx 0
L
0
T d x i + Wb
2
dt b2 + · · · ,
2
0
(3.3.4)
0
where the ellipsis denotes data penalties.
Exercise 3.3.1 Derive the following Euler–Lagrange equations for extrema of (3.3.4): −
∂λ ∂λ d Uˆ ∂λ − Uˆ = uˆ + (· · ·), ∂t ∂x du ∂ x λ=0
at t = T ,
(3.3.5) (3.3.6)
%
& ˆ d U Uˆ + uˆ λ = 0 du
(3.3.7)
at x = L, ∂ uˆ ∂ ˆ ˆ = F + W −1 + {U u} f λ, ∂t ∂x uˆ = I + Wi−1 λ
(3.3.8) (3.3.9)
76
3. Implementation
at t = 0, and
% uˆ = B +
& ˆ d U Uˆ + uˆ λ du
Wb−1
(3.3.10)
ˆ Note the “smoothness” assumption: U is differentiable with at x = 0; Uˆ ≡ U (u). respect to u. In (3.3.5), (· · ·) denotes data impulses. 3.3.3
Iteration schemes
The system (3.3.5)–(3.3.10) is nonlinear, and so representers are of no immediate use. All kinds of iteration schemes suggest themselves, but the following two schemes have met with success:
Scheme A ∂λn ∂λn − − Uˆ n−1 = uˆ n−1 ∂t ∂x
d Uˆ du
∂λn−1 + (· · ·)n , ∂x
λn = 0 at t = T ,
%
(3.3.12)
Uˆ n−1 + uˆ n−1
(3.3.11)
n−1
d Uˆ du
&
λn = 0
(3.3.13)
n−1
at x = L, ∂ uˆ n ∂ ˆ + U n−1 uˆ n = F + W −1 f λn , ∂t ∂x uˆ n = I + Wi−1 λn at t = 0, and
% uˆ n = B + Wb−1 Uˆ n−1 + uˆ n−1
d Uˆ du
(3.3.14) (3.3.15)
& λn
(3.3.16)
n−1
at x = 0. The system (3.3.11)–(3.3.16) is linear in uˆ n and λn ; it constitutes the Euler– Lagrange equations for a linear leastsquares problem, and may be solved either with representers (Bennett and Thorburn, 1992) or with the sweep algorithm of §4.2. Proving convergence of the sequence {uˆ n , λn }∞ n=1 to a solution of (3.3.5)–(3.3.10) is most difﬁcult in general, and has almost never been accomplished. Nevertheless, the sequence often seems to converge in practice, although the “source term” on the righthand side of (3.3.11) may need spatial smoothing. If it is smoothed, uˆ doesn’t quite minimize J . However, the approximate uˆ sufﬁces if Jˆ is less than the expected value M, which is the number of data. The righthand sides of (3.3.5) and (3.3.10) can cause difﬁculties because the calculus of the ﬁrst variation involves a linearization of the dynamics, much as in a linear stability analysis. Speciﬁcally, the adjoint dynamics
3.3 Nonlinear and nonsmooth estimation
77
of (3.3.5) and its iterate (3.3.11) involve advective coupling, respectively, of λ and ∞ ˆ ˆ {λn }∞ n=1 to the reference ﬂow, respectively U and {U n−1 }n=1 , thus the adjoint dynamics can be destabilized. Note that (3.3.5) and (3.3.11) lack the potential for amplitude modulation that is present in the nonlinear forward dynamics (3.3.1).
Scheme B
% ˆ d U ∂λn − Uˆ n−1 + uˆ n−1 − ∂t du
&
n−1
∂λn = (· · ·)n , ∂x λn = 0
at t = T ,
%
Uˆ n−1 + uˆ n−1 at x = L, ∂ uˆ n ∂ + ∂t ∂x
d Uˆ du
(3.3.17) (3.3.18)
&
λn = 0
(3.3.19)
n−1
d Uˆ n−1 Uˆ n−1 uˆ n + (uˆ n − uˆ n−1 )uˆ n−1 du
= F + W −1 f λn ,
uˆ n = I + Wi−1 λn at t = 0,
% uˆ n = B +
Wb−1
Uˆ n−1 + uˆ n−1
(3.3.20) (3.3.21)
d Uˆ du
& λn
(3.3.22)
n−1
at x = 0. This scheme, due to H.E. Ngodock, employs the tangent linearization of (3.3.1) (Lions, 1971; Le Dimet and Talagrand, 1986). The linearized momentum equation (3.3.20) yields the Euler–Lagrange equation (3.3.17). The latter has no inhomogeneity on the rhs other than the usual impulses proportional to the data misﬁts of uˆ n . The inhomogeneous term that appears on the rhs of (3.3.11) is now on the lhs of (3.3.17). That is, the term has become part of the adjoint operator. Furthermore, Scheme B would seem additionally risky, as it may introduce further linear instability into the forward dynamics (3.3.20). Scheme B obviates the need to compute and store a “ﬁrstguess” adjoint ﬁeld λ Fn that is the response to the “source term” in (3.3.11). Note 1. There are many heuristic iteration schemes such as A. There is only one tangent linearization scheme B; it follows from the series expansion of the nonlinear ﬂux in (3.3.1): dU U (u n )u n = U (u n−1 )u n−1 + (u n−1 ) (u n − u n−1 )u n−1 du + U (u n−1 )(u n − u n−1 ) + · · · (3.3.23) dU = (u n − u n−1 )u n−1 + Un−1 u n + · · · (3.3.24) du n−1 as appears in (3.3.20).
78
3. Implementation
Note 2. If sequences of equations (3.3.11)–(3.3.16) or (3.3.17)–(3.3.22) are solved using representers directly, then the latter must be recomputed for each iterate (value of n). Alternatively, the iterative, indirect construction of the representer solution must be repeated, for each such “linearizing” or “outer” iterate (value of n). In principle, the indirect approach would require a recalculation of the preconditioner for each outer iterate, but in practice such effort does not seem necessary for n > 2. Note 3. Since U is a smooth function of u, we can calculate the gradient of J with respect to u(x, t); for example, δJ ∂λ dU ∂λ = −2 + u +U , (3.3.25) δu(x, t) ∂t du ∂x where
λ ≡ Wf
∂u ∂ + {U u} − F , ∂t ∂x
(3.3.26)
if 0 < x < L, 0 < t < T and (x, t) is not a data point. (Readers unfamiliar with functional differentiation as in (3.3.25) may prefer to revisit this section after studying the discrete analog in §4.1.) Thus, given the ﬁeld u = u(x, t) we can evaluate λ(x, t) and hence the gradient of J , enabling a gradient search for the ˆ ﬁeld uˆ = u(x, t) that satisﬁes δJ ˆ = 0. [u] δu
(3.3.27)
Only one level of iteration is needed for this “state space” search: there is no need for two levels as in the doubly iterated representer approach or “data space search”. However, preconditioning is still essential for a state space search; in effect the inverse of the Hessian form H≡
δ2J δu(x, t)δu(y, s)
(3.3.28)
is required. Calculating H in full is usually prohibitive, as is inverting H . Some approximations, such as replacing H with its diagonal, do seem useful. See also §4.1.5. 3.3.4
Real dynamics: pitfalls of iterating
The idealized nonlinear wave dynamics of (3.3.1) provide a conveniently simple setting for the introduction of iterative solution schemes. The linear dynamics of Scheme A, as displayed in (3.3.14), retain the character of those in (3.3.1). However, the linear dynamics of Scheme B as shown in (3.3.20) are, as already indicated, of a different character. This can have radical consequences for real dynamics.
3.3 Nonlinear and nonsmooth estimation
79
(i) Continuity Consider ﬁrst an equation for conservation of volume, as appears in shallowwater models, layered models (Bleck and Smith, 1990) or indeed any reducedgravity Primitive Equation model (e.g., Gent and Cane, 1989): ∂h + u · ∇h + h∇ · u = 0, ∂t
(3.3.29)
where x = (x, y), ∇ = ( ∂∂x , ∂∂y ), u = (u, v), h = h(x, t) and u = u(x, t). Deﬁning X(at) to be the position at time t of a ﬂuid particle that was initially at position a, that is, dX (at) = u(X, t), dt
(3.3.30)
X(a0) = a,
(3.3.31)
subject to
and deﬁning h(at) and u(at) by h(at) ≡ h(X(at), t),
u(at) ≡ u(X(at), t)
(3.3.32)
allows us to express (3.3.29) as Dh (at) + h(at)(∇ · u)(at) = 0, Dt where the Lagrangian derivative is Dh ∂h . (at) ≡ (x, t) + u(x, t) · ∇h(x, t) Dt ∂t x=X(at)
(3.3.33)
(3.3.34)
The formal solution of (3.3.33) is
t h(at) = h(a0) exp − (∇ · u)(as) ds ,
(3.3.35)
0
which, so long as ∇ · u remains integrable in time, cannot change sign. The ocean cannot “dry out”, nor can ocean layers “outcrop” in a ﬁnite time. With smallamplitude gravity waves in mind, it is tempting to apply Scheme A to (3.3.29) as follows: ∂h n + un−1 · ∇h n = −h n−1 ∇ · un . ∂t
(3.3.36)
Together with a matching linearization of the momentum equations, (3.3.36) would capture such waves. However, the relegation of the divergence to a source term, as far as (3.3.36) alone is concerned, may cause h n to change sign just as though it were the perturbation amplitude of a small wave. Alternatively,
80
3. Implementation
applying Scheme A as follows: ∂h n + un−1 · ∇h n + h n ∇ · un−1 = 0, ∂t
(3.3.37)
preserves the positivity of h n . Scheme B leads uniquely to ∂h n + un−1 · ∇h n + h n ∇ · un−1 ∂t = −(un − un−1 ) · ∇h n−1 − h n−1 ∇ · (un − un−1 ),
(3.3.38)
which does not ensure positivity of h n . (ii) Thermodynamics Consider the turbulent transfer of heat, modeled simply by ∂T ∂ ∂T = K , ∂t ∂z ∂z
(3.3.39)
where the positive eddy conductivity K is a function of the temperature gradient: ∂T K =K > 0. (3.3.40) ∂z It is commonly assumed that K is a function of the gradient Richardson number (see, for example, Pacanowski and Philander, 1981), but it sufﬁces for this discussion to consider just (3.3.40). Linearizing (3.3.39) with Scheme A leads naturally to ∂ Tn ∂ Tn ∂ = K n−1 , (3.3.41) ∂t ∂z ∂z where K n−1 ≡ K (∂ Tn−1 /∂z) > 0. Both (3.3.39) and (3.3.41) yield wellposed initialvalue problems for t > 0. Scheme B leads uniquely to ∂ Tn ∂ = ∂t ∂z
K n−1 + K n−1
∂ Tn−1 ∂z
∂ Tn ∂z
− K n−1
∂ Tn−1 ∂z
2 ,
(3.3.42)
where K (θ) = d K (θ )/dθ , which is commonly assumed to be negative, for ∂ Tn−1 /∂z) to change sign in the θ > 0. Indeed, it is possible for (K n−1 + K n−1 tropical Paciﬁc Ocean, rendering (3.3.42) illposed for forward integration. The associated Euler–Lagrange equation would therefore be illposed for backward integration. The preceding examples show that while tangent linearization (“Scheme B”) has the merits of unique deﬁnition and efﬁcient implementation, it can in principle lead to unrealistic dynamics. If variational assimilation with a realistic model leads to difﬁculties, it is advisable to experiment with the linearization scheme.
3.3 Nonlinear and nonsmooth estimation
3.3.5
81
Dynamical linearization is not statistical linearization
Consider again the nonlinear wave equation (3.3.1), subject to some initial and boundary conditions that need not be considered here explicitly. The prior solution u F obeys ∂u F ∂ {U (u F )u F } = F. + ∂t ∂x
(3.3.43)
Regarding the solution u of (3.3.1) as the true circulation, the error in u F is v = u − u F . It follows that ∂v ∂ {U (u F + v)(u F + v) − U (u F )u F } = f. + ∂t ∂x
(3.3.44)
It has been assumed in the analyses of preceding chapters that E f = 0, that is, F is an unbiased estimate of the true forcing F + f . (We may always assume that the hypothetical ﬁeld E f vanishes, as a nonvanishing ﬁeld may be absorbed into F. The hypothesized mean, vanishing or nonvanishing, may of course be wrong.) If U were constant, then (3.3.44) would become ∂v ∂v +U = f. ∂t ∂x
(3.3.45)
Hence, as the expectation E is a linear operator: ∂(Ev) ∂(Ev) +U = 0, ∂t ∂x
(3.3.46)
for which the solution is Ev = 0, subject to suitable (linear, unbiased) initial and boundary conditions. In the general case U depends upon u, and f is not identically zero, thus it cannot be concluded that Ev = 0. The statistical variability of the forcing f can induce a bias in the circulation u, even though the prior estimate of forcing is unbiased. Moreover, closed forms such as (2.2.8) are no longer available for the covariance of u. Now consider a simple iteration of (3.3.1) about the iterate uˆ n−1 for an inverse ˆ estimate u: ∂u ∂ {U (uˆ n−1 )u} = F + f. + ∂t ∂x
(3.3.47)
The prior solution u Fn satisﬁes ∂u Fn ∂ + {U (uˆ n−1 )u Fn } = F, ∂t ∂x
(3.3.48)
hence the prior error vn = u − u Fn satisﬁes ∂vn ∂ {U (uˆ n−1 )vn } = f. + ∂t ∂x
(3.3.49)
The operator in (3.3.49) is dynamically linear. A linear superposition of forcing yields a linear superposition of solutions. That is the case, as long as the dependence of U (uˆ n−1 ) upon the actual forcing error f is ignored. In fact uˆ n−1 is a linear combination
3. Implementation
82
of representers with coefﬁcients proportional to the prior data misﬁt. The latter is of the form d − L[u Fn−1 ],
(3.3.50)
where d is the data. Of course, u Fn−1 is dependent on f via uˆ n−2 , and so on, but it sufﬁces here to note just that d = L[u] + ,
(3.3.51)
where u satisﬁes (3.3.1) and is the vector of measurement errors. The forcing error f that appears in the nonlinear model (3.3.1) is the same as the forcing error f that appears in the iterated error model (3.3.49). In particular, E {U (uˆ n−1 )vn } = U (uˆ n−1 )Evn ,
(3.3.52)
and so we cannot conclude that Evn vanishes identically when E f does. The dynamically linear model (3.3.49) is statistically nonlinear. Signiﬁcance tests and posterior error covariances for inverses of (3.3.47), without regard to its statistical nonlinearity, are at best guides rather than rigorous results. A strong warning from these guides, to the effect that the hypothesized prior means and covariances for f and for the initial and boundary inputs are unreliable, should nevertheless be taken seriously.
3.3.6
Linear, smooth dynamics; nonleastsquares
Suppose that the model is our original linear wave equation (1.2.6) and ancillary information (1.2.7)–(1.2.8), but suppose that our estimator or penalty functional is, for whatever reason, quartic in the residuals: T Q[u] = K f
L d x f + Kb
dt 0
T
0
L dt b + K i
4
dx i + k
4
0
4
M
m4 .
The gradient of Q with respect to u(x, t) is still welldeﬁned; for example, δQ ∂µ ∂µ = −4 +c , δu(x, t) ∂t ∂x where
µ = Kf
∂u ∂u +c −F ∂t ∂x
(3.3.53)
m=1
0
(3.3.54)
3 ,
(3.3.55)
if 0 < x < L, 0 < t < T and (x, t) is not a data point. Of course, we could ex1 press (3.3.54) in terms of µ 3 . A gradientsearch in state space is in principle indifferent to the nonlinearity of (3.3.53) and (3.3.54), but the nonlinearity of the Euler–Lagrange equations for nonquadratic penalties precludes the use of representers without a linearizing iteration. Attempts to do so have not been reported in a meteorological or oceanographic context.
3.3 Nonlinear and nonsmooth estimation
83
Exercise 3.3.2 Derive the Euler–Lagrange equations for extrema of Q deﬁned by (3.3.53).
3.3.7
Nonsmooth dynamics, smooth estimator
Suppose that the phase velocity in our ﬁrstorder wave equation is a nonsmooth function of the circulation, for example ∂u ∂ + {U (u)u} = F + f, ∂t ∂x
(3.3.56)
where U (u) =
u 0
u>0 u < 0.
(3.3.57)
This type of continuous but nondifferentiable parameterization of convection is especially common in mixedlayer models, and in models of convective adjustment. See for example Zebiak and Cane (1987) and Cox and Bryan (1984), respectively. The circulation variable u is usually temperature, while the “mixingfunction” U depends upon the vertical velocity, which is related to temperature via the dynamics. The difﬁculty is that the ﬁrst variation of f with respect to u is not deﬁned at u = 0. The situation may be handled using engineering “optimal control theory”, but that seems misguided in this context. There really are nonsmooth dependencies in engineering, but in geophysical ﬂuid dynamics we are merely making a crude parameterization of unresolved processes. Consider a ﬁnitedifference approximation to (3.3.56). The values of u at grid points are representatives of values within intervals. Yet, in reality, there will be a range of values within an interval, and convection will not start everywhere simultaneously. It therefore seems more sensible to replace a nonsmooth dependence as in (3.3.57) with a smooth dependence such as U (u) = ln (1 + eu/ ),
(3.3.58)
where is small. Clearly, U ∼u U = ln 2 and U ∼ eu/ ∼ 0
as u → ∞, at u = 0, as u → −∞.
(3.3.59)
Then dU eu/ = , du 1 + eu/
(3.3.60)
84
3. Implementation
so dU ∼ 1 as u → ∞, du dU = 1 at u = 0, 2 du and dU ∼ eu/ ∼ 0 as u → −∞. du In particular,
dU du
(3.3.61)
< 1 for all u. Had there been a discontinuity in U at u = 0, say U=
u>0 u < 0,
c 0
(3.3.62)
then, while it would be possible to smooth over the discontinuity in some small interval around u = 0, the derivative dU would be very large in that interval. In any case (3.3.62) du would seem an implausible parameterization of a gfd process. Even the mildest departure from nonsmoothness can greatly complicate an otherwise simple variational problem. For a lengthy investigation, see Xu and Gao (1999) and the references therein. Miller, Zaron, and Bennett (1994) consider varying the times of onset and end of convective mixing in a trivial linear model. The resulting variational problem is highly nonlinear, and the solution is not unique unless there are penalties for errors in prior estimates of the timing of convection. More recently, Zhang et al. (2000) ﬁnd that imposing smoothness on even a trivial model by “ﬁddling” may do more computational harm than good. They also explore the concept of subgradients of nonsmooth functionals, and examine the resulting generalizations of gradient searches in state space.
3.3.8
Nonsmooth estimator
Now suppose that the estimator for (3.3.56) is not smooth, such as: T J [u] =
L dx  f  + · · · ,
(3.3.63)
d x sgn{ f 0 −  f },
(3.3.64)
x >0 x < 0.
(3.3.65)
dt 0
0
or T J [u] =
L dt
0
0
where sgn(x) =
1 −1
3.3 Nonlinear and nonsmooth estimation
85
Neither of these functionals has uniformly welldeﬁned ﬁrst variations with respect to u. Consequently there is no gradient information that can guide a search, and Euler– Lagrange conditions cannot be formulated, unless we are sure of special information such as the sign of fˆ. Only bruteforce minimization is available in the absence of gradient information, but brute force can be applied thoughtfully: see the method of simulated annealing in §4.4.
Chapter 4 The varieties of linear and nonlinear estimation
A data space search is the most efﬁcient way to solve a linear, leastsquares smoothing problem deﬁned over a ﬁxed time interval. The method exploits linearity, and so is unavailable for nonlinear dynamics, or for penalties other than leastsquares. As discussed in Chapter 3, a data space search may be conducted on linear iterates of the nonlinear Euler–Lagrange equations. The existence of the nonlinear equations implies that the penalty is a smooth functional of the state, in which case a state space search may always be initiated. The nature of state space searches is intuitively clear, and their use is widespread. Conditioning degrades as the size of the state space gets very large. Collapsing the size of the state space by assuming “perfect” dynamics is the basis of “the” variational adjoint method: only initial values, boundary values and parameter values are varied. Preconditioning may in principle be effected by use of secondorder variational equations, but even iterative construction of the state space preconditioner is unfeasible for highly realistic problems. Technique for numerical integration of variational equations is not a paramount consideration, but deserving of attention since it can be consuming of human time. Operational forecasting is inherently sequential; data are constantly arriving and forecasts must be issued regularly. In such an environment, it is more natural to ﬁlter a model and data sequentially than to smooth them over a ﬁxed interval. The Kalman ﬁlter is available for linear dynamics; it forecasts the covariance of the error in the forecast of the state, and uses the covariance to make a leastsquares spatial interpolation of newly arriving data. If the dynamics are nonlinear, the Kalman ﬁlter may be applied iteratively. It is computationally expensive, even in the linear case, and various economies are practised.
86
4.1 State space searches
87
Leastsquares is the estimator of maximum likelihood for Gaussian or normal random quantities. It may be applied to the estimation of any quantity, but the reliability of the estimate would be in doubt. Maximum likelihood estimators are deﬁned for any random variable, yet are subtle or even ambiguous for even the simplest of nonnormal distributions. Monte Carlo methods for simulating random quantities have great appeal, owing to their conceptual simplicity. Ingenious algorithms exist for generating random variables of any distribution. The same algorithms lead to equally ingenious optimization methods for arbitrary penalties.
4.1
State space searches
4.1.1
Gradients
The representer algorithm described in the previous chapters is an arcane and intricate method for minimizing quadratic penalty functionals such as J in (1.2.9). It would be much simpler to evaluate the gradient of J with respect to changes in the errors f, i and b, and then use the gradient information in a search for the minimum of J . Na¨ıve evaluation of the gradient would be prohibitively expensive, but an economical alternative exists. We may avoid the abstraction of a gradient in function space, by referring to numerical models. Then the gradient becomes a ﬁnitedimensional vector with as many components as there are independently variable quantities, or “controls”. These may be as numerous as the entire set of gridded variables, in which case the search would be hopelessly illconditioned. Imposing the dynamics as a strong constraint reduces the set of controls to the initial conditions, boundary conditions and dynamical parameters such as diffusivities. Even so, these may be numerous and convergence may be slow. Crude preconditioners are unreliable, while full preconditioning remains unfeasible. The section concludes with some hints on deriving gradients. 4.1.2
Discrete penalty functional for a ﬁnite difference model; the gradient
For simplicity, the following discussion is based on an “ocean model” of maximal simplicity. It is the ordinary differential equation du (t) = F(t) + f (t), dt
(4.1.1)
u(0) = I + i.
(4.1.2)
where 0 ≤ t ≤ T , subject to
88
4. The varieties of linear and nonlinear estimation
In (4.1.1) and (4.1.2), F and I denote prior estimates of forcing and initial values, while f and i denote errors in those estimates. It sufﬁces to consider a single datum: u(td ) = U + ,
(4.1.3)
where U is the datum at time td , and is the measurement error. A leastsquares estimator for f , i, and hence u is T J [u] = W f
dt f 2 + Wi i 2 + w 2
(4.1.4)
0
T = Wf
dt
du −F dt
2 + Wi (u(0) − I )2 + w(u(td ) − U )2 .
(4.1.5)
0
We may avoid functional analysis by replacing the integrals and derivatives in (4.1.5) with sums and differences. To this end, deﬁne tn = nt
(4.1.6)
t = T /N ;
(4.1.7)
for 0 ≤ n ≤ N , where
hence t0 = 0 ,
t N = T.
(4.1.8)
Assume td = n d t
(4.1.9)
0 < nd < N .
(4.1.10)
u n ≡ u(tn )
(4.1.11)
for some integer n d :
Deﬁne
for 0 ≤ n ≤ N , and approximate derivatives with forward differences: du u n+1 − u n . dt n t
(4.1.12)
Replacing the integral with a sum, (4.1.5) becomes JN = W f (t)−1
N −1
(u n+1 − u n − t Fn )2 + Wi (u 0 − I )2 + w(u d − U )2 ,
(4.1.13)
n=0
where Fn ≡ F(tn )
and u d ≡ u(td ) = u n d .
(4.1.14)
4.1 State space searches
89
Thus the N point approximation to J [u] is JN = JN (u 0 , u 1 , . . . , u N ) = JN (u),
(4.1.15)
where u = (u 0 , . . . , u N ) R N .
(4.1.16)
There are very sophisticated ways (Press et al., 1986) to use the information in ∇JN in order to minimize JN , but we only need consider the most elementary for ˆ and the latest is uk . If now. Suppose we have made k estimates of the minimizer u, k k JN ≡ JN (u ) is not satisfactorily small, we may na¨ıvely improve on uk as follows: uk+1 = uk − θ∇JNk for some small positive number θ. Then of course JNk+1 ≡ JN (uk+1 ) = JN uk − θ∇JNk 2 ∼ = JN (uk ) − θ ∇JNk < JN (uk ).
(4.1.17)
(4.1.18) (4.1.19) (4.1.20)
Now
∇JNk
n
∂ JN k (u ) ∂u n JN u k0 , u k1 , . . . , u kn + u n , . . . , u kN − JN (uk ) ∼ , = u n
≡
(4.1.21)
where u n is a small increment in u n , and so we may evaluate (4.1.21) by direct substitution into (4.1.13). 4.1.3
The gradient from the adjoint operator
Evaluation of ∇JN via (4.1.21) requires N evaluations of JN , for each of N increments u n , 1 ≤ n ≤ N . As an alternative, we may evaluate the gradient more efﬁciently by ﬁrst doing some elementary calculus. It follows easily from (4.1.13) that ∂ JN u1 − u0 = −2W f (4.1.22) − F0 + 2Wi (u 0 − I ), ∂u 0 t ∂ JN u n − u n−1 u n+1 − u n = 2W f − Fn−1 − + Fn ∂u n t t + 2wδnn d (u d − U ) for 0 < n < N , and ∂ JN = 2W f ∂u N
u N − u N −1 − FN −1 . t
(4.1.23)
(4.1.24)
90
4. The varieties of linear and nonlinear estimation
Hence ∇JN may be evaluated at roughly the same cost as one evaluation of JN . Now deﬁne the weighted dynamical residual: u n+1 − u n λn ≡ W f − Fn (4.1.25) t for 0 ≤ n ≤ N − 1. Then (4.1.22)–(4.1.24) become 1 ∂ JN 2 ∂u 0 = −λ0 + Wi (u 0 − I ) , (2t)−1
δnn ∂ JN λn − λn−1 + w d (u d − U ) =− ∂u n t t
(4.1.26) (4.1.27)
for 0 < n < N − 1, and 1 ∂ JN 2 ∂u N = λ N −1 .
(4.1.28)
Exercise 4.1.1 Derive the Euler–Lagrange equations for the penalty functional (4.1.5). Discretize as in (4.1.6)–(4.1.12). Compare with (4.1.25)–(4.1.28). The discrete Euler–Lagrange equations are simply ∇JN = 0.
(4.1.29)
In the continuous limit, we might write this as δ J [u] = 0. δu(x, t)
(4.1.30)
So our procedure is now: (i) (ii) (iii) (iv)
4.1.4
estimate uk ; evaluate λk by substitution of uk into (4.1.25); evaluate ∇JNk by substitution of λk into (4.1.26)–(4.1.28); and then estimate uk+1 using the gradient information, as in (4.1.17) for example.
“The” variational adjoint method for strong dynamics
There is a special case that has been very widely used (“THE variational adjoint method”, e.g., Lewis and Derber, 1985; LeDimet and Talagrand, 1986; Thacker and Long, 1988; Greiner and Perigaud, 1994a,b; Kleeman et al., 1995). Assume that the dynamics are perfect; that is, let W f /Wi , W f /w → ∞.
(4.1.31)
4.1 State space searches
91
Set ∂ JNk = 0 for 0 < n < N − 1, ∂u n
(4.1.32)
∂ JNk = 0. ∂u N
(4.1.33)
and set
Then λkn may be determined for 0 ≤ n ≤ N − 1, by solving (4.1.32) and (4.1.33), that is, by stepping λkn−1 − λkn δnn = −w d u kd − U t t
(4.1.34)
λkN −1 = 0.
(4.1.35)
backwards from
∂ Jk
may be estimated This leads to a value for λk0 , and hence for ∂uN0 via (4.1.26). Then u k+1 0 using this remaining amount of nonvanishing gradient information: u k+1 = u k0 − θ 0
∂ JNk . ∂u 0
(4.1.36)
The rest of uk+1 is evaluated using (4.1.25): k+1 u k+1 n+1 − u n k = Fn + W −1 f λn , t
(4.1.37)
for 0 ≤ n < N − 1. But λn is O(Wi , w), and so W −1 f λn → 0 in the limit as W f /Wi → ∞, W f /w → ∞. In other words, each estimate of uk obeys the exact or “strong” dynamical constraint u kn+1 − u kn = Fn t
(4.1.38)
for 0 ≤ n ≤ N − 1. Notice that only the initial value u k0 is varied independently. The values of u k1 , . . . , u kN also vary, but they are determined by the value of u k0 and by (4.1.38). That is, a forward integration is required. A backward integration is also required in order to determine λk . In practice we deal with partial differential equations, and the initial value u k0 is a ﬁeld: u k0 = u k0 (x). If the spatial grid is ﬁne, then the initial ﬁeld u k0 becomes a long vector, while the timedependent solution or “state” (u k1 (x), . . . , u kN (x)) becomes such a long vector that convergence of descent algorithms is typically very slow. Thus the “weakconstraint” minimization described in §4.1.2 is quite unfeasible, while the “strongconstraint” minimization would seem feasible. Note the imposition of a major physical constraint (exact or “strong” dynamics), in order to deal with a computational difﬁculty. We may well be able to ﬁt the data fairly closely with an exact solution of a model, but this is of dubious value when, as usual, we have more conﬁdence in the data
92
4. The varieties of linear and nonlinear estimation
than in the model. We should instead make estimates of the dynamical error variances and hence the dynamical weight W f , and then use a minimization technique that is adequate to the task. Furthermore, a reasonable hypothesis is then being tested.
4.1.5
Preconditioning: the Hessian
Convergence towards the minimum of J N may be accelerated by the use of secondorder information. The Taylor expansion of JN (u) about the minimum uˆ is 1 ˆ − u) ˆ T H(u ˆ + ···, ˆ + (u − u) ˆ T ∇JN (u) ˆ + (u − u) JN (u) ∼ = JN (u) 2
(4.1.39)
ˆ are where the components of the Hessian matrix H 2 ˆ nm = ∂ JN (u), ˆ H ∂u n ∂u m
(4.1.40)
for 1 ≤ n, m ≤ N . Since uˆ is an extremum of JN , ˆ = 0, ∇JN (u)
(4.1.41)
1 ˆ − u). ˆ ˆ + (u − u) ˆ T H(u JN (u) ∼ = JN (u) 2
(4.1.42)
hence
ˆ the gradient of JN is approximately given by So, near u, ˆ − u). ˆ ∇JN (u) ∼ = H(u
(4.1.43)
ˆ −1 ∇JN (u) ∼ ˆ H = u − u.
(4.1.44)
ˆ is invertible, Assuming that H
If we step in the direction of (4.1.44), then ˆ uk+1 = uk − θ (Hk )−1 ∇JNk ∼ = uk − θ (uk − u).
(4.1.45)
The advantage of this approach is clear once we examine Fig. 2.1.1. If the Hessian is poorly conditioned, that is, if its eigenvalues have a large range, then contours of constant JN will be highly eccentric ellipsoids. A step from uk down the gradient ∇JNk ˆ However, a step from uk in the direction will not be a step towards the minimum at u. k −1 k ˆ ˆ and so convergence towards uˆ (H ) ∇JN will be a step approximately towards u, should be more rapid. It would seem that such “preconditioning” would be unfeasible if N is very large, since it would be so expensive to compute and invert the N × N matrix H at each step. However (Le Dimet et al., 1997), H−1 ∇J may be evaluated iteratively without ﬁrst computing H, just as P−1 h may be evaluated iteratively without ﬁrst computing P (see §3.1.4). A ﬁnal caution is that H is N × N , where N is the number of state variables, while P is M × M, where M is the number of data.
4.1 State space searches
93
Note. Inspecting (2.1.14) and (2.1.17) shows that −1 S ≡ R + R C−1 R = P C P − P,
(4.1.46)
where Si j =
∂ 2J . ∂βi ∂β j
(4.1.47)
Thus R and P are closely related to S, the Hessian of the penalty functional J [u] with respect to the observable degrees of freedom. For a review of the use of “adjoint models” see Errico (1997). 4.1.6
Continuous adjoints or discrete adjoints?
The derivation of gradients and Euler–Lagrange equations for penalty functions was illustrated in §4.1.2 with a trivial example. (Note that JN deﬁned by (4.1.13) is a realvalued function of the N dimensional vector u = (u 1 , . . . , u N )T , whereas J deﬁned by (4.1.5) is a realvalued functional of the function u = u(t).) Derivation of “discrete adjoints” becomes exacting and tedious as the complexity of the forward numerical model increases. Is it worth the trouble? After all, one could with comparative ease derive the adjoint operators analytically and then approximate numerically as in the forward model. In general, proceeding in that order “breaks” adjoint symmetry; the “discrete adjoint equation” is not the “adjoint discrete equation”. The broken symmetry manifests itself directly as a spurious, asymmetric part in the representer matrix. So long as the asymmetric part is relatively small, it could be discarded. The resulting inverse solution would be slightly suboptimal. In the indirect representer algorithm of §3.1.4, the representer matrix is not being explicitly constructed, hence its asymmetry cannot be suppressed. A preconditioned biconjugategradient solver must be used in the iterative search for the representer coefﬁcients. It seems cleaner to work with the “adjoint discrete equation”; then asymmetry of the representer matrix becomes a very useful indicator of coding errors. Again, deriving the adjoint discrete equation is an exacting task. Experience and technique are important. Recognition of pattern can greatly reduce the burden. As a rule, centered ﬁnitedifference operators are selfadjoint. Most difﬁculties occur at boundaries, where operators are typically onesided and so not selfadjoint. The introduction of virtual state variables outside the domain of the forward model can simplify the resulting adjoint discrete equation (J. Muccino, personal communication). For example, consider the simple conduction problem ∂T ∂2T (4.1.48) =θ 2 ∂t ∂x for constant θ > 0, for 0 ≤ x ≤ xmax and 0 ≤ t ≤ tmax , subject to the initial condition T (x, 0) = I (x)
(4.1.49)
94
4. The varieties of linear and nonlinear estimation
for 0 ≤ x ≤ xmax , and the heatreservoir boundary conditions T (0, t) = L(t),
T (xmax , t) = R(t)
(4.1.50)
for 0 ≤ t ≤ tmax . A simple, explicit timestepping scheme is provided by k k Tmk+1 − Tmk = (θ t/x 2 ) Tm+1 + Tm−1 − 2Tmk ,
(4.1.51)
where Tmk = T (mx, kt),
0 ≤ m ≤ M,
0 ≤ k ≤ K,
x = xmax /M,
t = tmax /K . (4.1.52)
It is straightforward to devise a penalty function for generalized inversion of the wellposed discrete forward problem corresponding to (4.1.48)–(4.1.50) plus an overdetermining set of temperature data. Inspection of (4.1.51) indicates that the spatial summation of weighted squares of residuals in (4.1.51) should range from m = 1 to m = M − 1. The resulting Euler–Lagrange equations are inelegant, but improve with the introduction of bogus weighted residuals λk0 and λkM that are identically zero. k k Alternatively, virtual temperatures T−1 and TM+1 may be introduced. Residuals λk0 k and λ M are now automatically deﬁned, and the spatial summations of squared residuals should be extended to 0 ≤ m ≤ M. The Euler–Lagrange equations are then far k k tidier. In particular, variation with respect to T−1 and TM+1 implies that λk0 and λkM vanish. Practitioners of spectral methods or ﬁnite element methods are free from all these vexed considerations. Derivatives are transferred to the basis functions or to the elements by partial integration, and are then evaluated analytically.
4.2
The sweep algorithm, sequential estimation and the Kalman ﬁlter
4.2.1
More trickery from control theory
The Euler–Lagrange equations are a twopoint boundaryvalue problem in the time interval of interest. We used representers in order to untangle this problem, that is, in order to express it in terms of initial value problems. The Gelfand and Fomin sweep algorithm provides a remarkable alternative (Gelfand and Fomin, 1963; Meditch, 1970). Partial implementation of the algorithm yields an appealing “sequential estimation” scheme for assimilating data. This controltheoretic derivation of the Kalman ﬁlter follows logically from preceding chapters, but the reader may prefer to start with the statistical derivation of the Kalman ﬁlter in §4.3. That is, §4.2 may be omitted from a ﬁrst reading.
4.2 Sweep algorithm
95
The sweep algorithm yields the Kalman ﬁlter
4.2.2
Let us gather up all the parts of the Euler–Lagrange system for our simple model (1.2.6)–(1.2.8) and for the penalty functional (1.5.7), in the most general form that we have used: −
∂λ ∂λ ˆ T C−1 (x, t) − c (x, t) = (d − u) δ, ∂t ∂x λ(x, T ) = 0,
(4.2.1) = (1.3.1) (4.2.2) = (1.3.2)
λ(L , t) = 0,
(4.2.3) = (1.3.3)
∂ uˆ ∂ uˆ (x, t) + c (x, t) = F(x, t) + (C f • λ)(x, t), ∂t ∂x ˆ u(x, 0) = I (x) + (Ci ◦ λ)(x), ˆ t) = B(t) + c(Cb ∗ λ)(t). u(0,
(4.2.4) = (1.5.14) (4.2.5) = (1.5.15) (4.2.6) = (1.5.16)
Recall that the covariances C f , Ci , Cb and C are the inverses of the weights W f , Wi , Wb and w: see §1.5. The impulse vector δ has, as its mth component, the scalar impulse δ(x − xm )δ(t − tm ). First, assume that the dynamical residual f is uncorrelated in time: C f (x, y, t, s) = Q f (x, y)δ(t − s),
(4.2.7)
and so the “forward” Euler–Lagrange equation (4.2.4) becomes ∂ uˆ ∂ uˆ (x, t) + c (x, t) = F(x, t) + ∂t ∂x
L dy Q f (x, y)λ(y, t).
(4.2.8)
0
Next, assume that the inverse estimate uˆ is linearly related to the weighted residual or adjoint variable λ: L ˆ u(x, t) =
dy P(x, y, t)λ(y, t) + v(x, t),
(4.2.9)
0
for some “slope” P and “intercept” v. With the substitution of (4.2.9), the lefthand side of (4.2.8) becomes
L dy 0
+
∂P ∂λ ∂P (x, y, t)λ(y, t) + P(x, y, t) (y, t) + c (x, y, t)λ(y, t) ∂t ∂t ∂x
∂v ∂v (x, t) + c (x, t). ∂t ∂x
(4.2.10)
Now assume that the measurement errors are uncorrelated, if at different times: n m = 0
if tn = tm
(4.2.11)
for 1 ≤ n, m ≤ M, in which case the M × M measurement error covariance matrix consists of K blocks on the diagonal, where each block is an N × N matrix; K is the
96
4. The varieties of linear and nonlinear estimation
number of measurement times, N is the number of measurement sites, and M = N K . That is, C = diag (C1 , . . . , CK ). Then “the” Euler–Lagrange equation (4.2.1) becomes −1 k ∂λ ∂λ (dk − uˆ k )T Ck δ , (x, t) − c (x, t) = ∂t ∂x k=1 K
−
(4.2.12)
where (dk )n is the datum at site xn at time tk , while (δ k )n = δ(x − xn )δ(t − tk ) and, according to the substitution (4.2.9), L k ˆ n , tk ) = dy P(xn , y, tk )λ(y, tk ) + v(xn , tk ). (uˆ )n ≡ u(x
(4.2.13)
(4.2.14)
0
Exercise 4.2.1 Now substitute (4.2.14) into the rhs of (4.2.12), and then (4.2.12) into (4.2.10), which is the lhs of (4.2.8). At times t = tk , 1 ≤ k ≤ K , integrate over y by parts and equate coefﬁcients of λ(y, t), yielding ∂P ∂P ∂P (x, y, t) + c (x, y, t) + c (x, y, t) ∂t ∂x ∂y + c P(x, 0, t)δ(y) = Q f (x, y),
(4.2.15)
and leaving ∂v ∂v (x, t) + c (x, t) = F(x, t). ∂t ∂x
(4.2.16)
From the initial condition (4.2.5), we may analogously obtain P(x, y, 0) = Ci (x, y)
(4.2.17)
v(x, 0) = I (x).
(4.2.18)
and
If we assume that the errors in the boundary data are uncorrelated in time: Cb (t, s) ≡ b(t)b(s) = Q b δ(t − s),
(4.2.19)
ˆ t) = B(t) + cQ b λ(0, t). u(0,
(4.2.20)
P(0, y, t) = cQ b δ(y)
(4.2.21)
v(0, t) = B(t).
(4.2.22)
then (4.2.6) becomes
We recover
and
4.2 Sweep algorithm
97
The system (4.2.15), (4.2.17), (4.2.21) and the system (4.2.16), (4.2.18), (4.2.22) may be integrated forward in time until t = t1 −. Note that the initial value and forcing for P(x, y, t), respectively Ci (x, y) and Q f (x, y), are symmetric. If we assume that P(x, y, t) is symmetric, then by (4.2.21) P(x, 0, t) = P(0, x, t) = cQ b δ(x),
(4.2.23)
and hence the seemingly symmetrybreaking fourth term on the lhs of (4.2.15) is c P(x, 0, t)δ(y) = c2 Q b δ(x)δ(y),
(4.2.24)
which is symmetric. In other words, assuming symmetry leads to no contradiction. It remains to determine the jumps in P and v as t passes through tk . First, we learn from (4.2.12) that λ is discontinuous in tk , with −1 −λ(x, tk +) + λ(x, tk −) = (dk − uˆ k )T Ck δ, (4.2.25) where (δ)n = δ(x − xn ). We infer from (4.2.4) that uˆ is continuous at tk . Hence (4.2.9) implies that L tk + dy P(x, y, t)λ(y, t) + v(x, t) = 0 . (4.2.26) tk −
0
Substitute (4.2.14) into (4.2.25) and then substitute (4.2.25) into (4.2.26). Equating terms proportional to λ(x, y, tk −) yields, after a little algebra, P(x, y, tk +) − P(x, y, tk −) −1 k− = − Pk− (x)T Pk− + Ck P (y)
(4.2.27)
and T
v(x, tk +) − v(x, tk −) = Kk (dk − vk− ),
(4.2.28)
Pnk− (x) ≡ P(xn , x, tk −)
(4.2.29)
where
k− Pnm ≡ P(xn , xm , tk −) −1 k+ Kk ≡ Ck P
(4.2.30) (4.2.31)
and (vk− )n = v(xn , tk −).
(4.2.32)
We now have an explicit algorithm for P and v, for all t ≥ 0. To complete the formula ˆ We may ˆ we need λ. Now λ obeys (4.2.1)–(4.2.3), but (4.2.1) involves u. (4.2.9) for u, eliminate uˆ using (4.2.14), yielding an equation for P, v and λ. We can determine P and v, so λ may be found by backwards integration, and the Gelfand and Fomin sweep is complete.
98
4. The varieties of linear and nonlinear estimation
Exercise 4.2.2 ˆ Derive the equation for λ, free of u.
Note 1. The above procedure has a major drawback: it would be necessary to compute and store P(x, y, t) and v(x, t) for 0 < x, y < L, and for 0 < t < T . This would be prohibitive in practice. Note 2. It is only necessary to store P(x, y, tk +) and v(x, tk +) in order to evaluate uˆ k , for 1 ≤ k ≤ K , and hence λ (see(4.2.12)). Having solved for λ, we could then ﬁnd uˆ by integrating (4.2.4)–(4.2.6). Note 3. There are other such “control theory” algorithms such as that of Rauch, Tung and Streibel (e.g., Gelb, 1974), but these require even more computation and storage. These algorithms are impractical for the generalized inversion of oceanic or atmospheric models. Note 4. The adjoint variable λ vanishes after assimilating the last data: λ = 0 for t K < t < T , hence the generalized inverse uˆ agrees with the “intercept” v at the end of the smoothing interval: ˆ u(x, t) = v(x, t)
(4.2.33)
for t K < t < T , where T is somewhat arbitrary. So, if we only want to know the inﬂuence of the K th (the latest) data d K upon the circulation estimate uˆ at time t K (the present), then we need not do more than solve for v (which requires solving for P: see (4.2.28)–(4.2.32)). The previous data: d1 , . . . , d K −1 also inﬂuence v at time t K , but d K has no inﬂuence on v for t < t K . Thus v is a “sequential” estimate of u, using data as they arrive.
4.3
The Kalman ﬁlter: statistical theory
4.3.1
Linear regression
The Kalman ﬁlter has just been derived as a ﬁrst step in solving linear Euler–Lagrange problems. It is a sequential algorithm, that is, it calculates the generalized inverse at ˆ times later than all the data: v(x, t) = u(x, t) for all t > t K . Recall that uˆ minimizes a quadratic penalty functional over 0 ≤ t ≤ T , where t K < T . The Kalman ﬁlter will now be derived using linear regression.
4.3.2
Random errors: ﬁrst and second moments
Our ocean model is ∂u ∂u +c = F + f, ∂t ∂x
(4.3.1)
4.3 The Kalman ﬁlter: statistical theory
99
for 0 ≤ x ≤ L and 0 ≤ t ≤ T , subject to the boundary condition u(0, t) = B(t) + b(t)
(4.3.2)
u(x, 0) = I (x) + i(x).
(4.3.3)
and the initial condition
We have assumed that F, B and I are unbiased estimates of the forcing, boundary and initial values: E f = Eb = Ei = 0,
(4.3.4)
and we prescribed the autocovariances of f , b and i: E( f (x, t) f (y, s)) = Q f (x, y)δ(t − s),
(4.3.5)
E(b(t)b(s)) = Q b δ(t − s),
(4.3.6)
E(i(x)i(y)) = Ci (x, y).
(4.3.7)
We assumed that their crosscovariances all vanish: E( f b) = E( f i) = E(bi) = 0. There are data at N points x1 , . . . , x N , at discrete times t = t1 , . . . , t K : dnk = u(xn , tk ) + nk
(4.3.8)
for 1 ≤ n ≤ N , where nk are the measurement errors, for which Ek = E( f k ) = E(bk ) = E(ik ) = 0, E( ) = k lT
δkl Ck
.
(4.3.9) (4.3.10)
That is, f , b and k are uncorrelated in time. The vectors in (4.3.9), (4.3.10) have N components. Note that the points x1 , . . . , x N do not necessarily coincide with a spatial grid for numerical integration of (4.3.1)–(4.3.3); they are merely a set of N measurement sites.
4.3.3
Best linear unbiased estimate: before data arrive
We shall now construct w(x, t), the best linear unbiased estimate of u(x, t), given data prior to t. Assuming t1 > 0, at time t = 0 we can do no better than w(x, 0) = I (x),
(4.3.11)
for which the error variance is Ci : see (4.3.7). For 0 ≤ t ≤ t1 −, let ∂w ∂w +c = F, ∂t ∂x
(4.3.12)
w(0, t) = B(t).
(4.3.13)
100
4. The varieties of linear and nonlinear estimation
The error e ≡ u − w obeys ∂e ∂e (x, t) + c (x, t) = f (x, t), ∂t ∂x e(x, 0) = i(x), e(0, t) = b(t),
the solution of which is t L e(x, t) =
(4.3.14) (4.3.15) (4.3.16)
t ds dξ γ (ξ, s, x, t) f (ξ, s) + c
0 0
ds γ (0, s, x, t)b(s) 0
L +
dξ γ (ξ, 0, x, t)i(ξ ),
(4.3.17)
0
where γ is the Green’s function (see §1.1.4). Hence E(e(x, t) f (y, t)) = 12 Q f (x, y), !t since γ (x, t, y, t) = δ(x − y), and t− δ(s) ds = 12 . Also E(e(x, t)b(t)) = cQ b δ(x).
(4.3.18)
Now deﬁne the spatial error covariance at time t by P(x, y, t) ≡ E(e(x, t)e(y, t)) = P(y, x, t).
(4.3.19)
Multiplying (4.3.14) by e(y, t) and averaging yields ∂P ∂P ∂P (x, y, t) + c (x, y, t) + c (x, y, t) = Q f (x, y); ∂t ∂x ∂y
(4.3.20)
multiplying (4.3.16) by e(y, t) and averaging yields P(0, y, t) = cQ b δ(y),
(4.3.21)
P(x, 0, t) = cQ b δ(x).
(4.3.22)
P(x, y, 0) = Ci (x, y).
(4.3.23)
Initially,
4.3.4
Best linear unbiased estimate: after data have arrived
1 The situation at time t1 − is that we have an estimate w− (x) ≡ w(x, t1 −), equal to the 1 mean of u(x, t1 ), and we have its error covariance P− (x, y) ≡ P(x, y, t1 −). The new information are the data d1 . These too contain random errors, but by (4.3.8) we are assuming that
Ed1 = Eu1 .
(4.3.24)
1 1 Let us seek a new estimate w+ (x) for u(x, t1 ) which is linear in w− (x) and associated data misﬁts: 1 1 w+ (x) = αw− (x) + s(x)T (d1 − w1− ),
(4.3.25)
4.3 The Kalman ﬁlter: statistical theory
101
where 1 (xn ) ≡ w(xn , t1 −). (w1− )n = w−
(4.3.26)
The constant α and the interpolant s(x) have yet to be chosen. Consider the error 1 1 e+ (x) = u(x, t1 ) − w+ (x).
(4.3.27)
1 1 u(x, t1 ) = w− (x) + e− (x),
(4.3.28)
Now 1 (x) = e(x, t1 −). Hence where e− 1 1 1 e+ (x) = (1 − α)w− (x) + e− (x) − s(x)T (d1 − w1− ).
(4.3.29)
1 But Ee− = 0. So if we choose α = 1, then 1 Ee+ (x) = 0
and (4.3.25) is an unbiased estimate. The error variance is 1 2 P+1 (x, x) = E e+ (x) .
(4.3.30)
(4.3.31)
Exercise 4.3.1 Show that the error variance (4.3.31) is least if the optimal interpolant is s(x) = K1 (x), where the “Kalman gain” vector ﬁeld K1 (x) in (4.3.32) is ' (−1 K1 (x) = P1− + C1 P1− (x).
(4.3.32)
(4.3.33)
The vector P1− (x) and matrix P1− have components 1 Pn− (x) = P−1 (xn , x), 1 Pnm− = P−1 (xn , xm ).
(4.3.34)
Exercise 4.3.2 Show that the posterior error covariance at time t1 is 1 1 P+1 (x, y) ≡ E e+ (x)e+ (y) ' (−1 = P−1 (x, y) − P1− (x)T P1− + C1 P1− (y).
(4.3.35)
Clearly, we may repeat this construction at t2 , t3 , . . . . See Fig. 4.3.1. Gathering up all the results, the Kalman ﬁlter estimate w satisﬁes ∂w ∂w +c =F ∂t ∂x
(4.3.36)
102
4. The varieties of linear and nonlinear estimation
Figure 4.3.1 Time line for the Kalman ﬁlter.
for 0 ≤ x ≤ L, tk < t < tk+1 , subject to w(x, 0) = I (x)
(4.3.37)
w(0, t) = B(t).
(4.3.38)
w(x, tk +) − w(x, tk −) = Kk (x)T (dk − wk− ),
(4.3.39)
and The change in w at time tk is
where the Kalman gain is (−1 ' Kk (x) = Pk− + Ck Pk− (x).
(4.3.40)
The error covariance satisﬁes ∂P ∂P ∂P +c +c = Qf ∂t ∂x ∂y
(4.3.41)
for 0 ≤ x, y ≤ L, tk < t < tk+1 , subject to P(0, y, t) = cQ b δ(y),
(4.3.42)
P(x, 0, t) = cQ b δ(x)
(4.3.43)
P(x, y, 0) = Ci (x, y).
(4.3.44)
P(x, y, tk +) − P(x, y, tk −) = − Pk− (x)T Kk (y).
(4.3.45)
and
The change in P at tk is
The new data always reduce the error variance at data sites. Note carefully the assumptions that the dynamical and boundary errors f , b and the data k are uncorrelated in time, and that the different types of errors are not crosscorrelated. Note also that the optimal choices for α and s(x) in (4.3.25) are not random. They depend not upon the 1 random inputs w− (x), d1 but upon the covariances of the errors in the inputs. 4.3.5
Strange asymptotics
It is usually assumed that the data errors are statistically stationary, that is, Ck is independent of k. It is often the case that the temporal sampling interval tk+1 − tk also is independent of k. Consequently, the Kalman ﬁlter error covariance P approaches an equilibrium state, in which P(x, y, tk −) = P(x, y, tk+1 −) and P(x, y, tk +) = P(x, y, tk+1 +). The
4.3 The Kalman ﬁlter: statistical theory
103
covariance does still evolve in time from tk + to tk+1 −, but Q f and Q b are independent of t and so the evolution is the same in every data interval. In general we are interested in more complicated dynamics than are expressed in (4.3.1); so long as the dynamics are linear, they may be expressed as ∂u (4.3.46) + L x u = F + f, ∂t where L x is a linear partial differential operator with respect to x. Of course, Primitive Equation models involve many dependent variables, but we shall retain just one here, namely u, for clarity. The error covariance now satisﬁes ∂P (4.3.47) (x, y, t) + L x P(x, y, t) + L y P(x, y, t) = Q f (x, y). ∂t To simplify the discussion further, let us assume that the data interval t = tk+1 − tk is much smaller than the evolution time scale for (4.3.47), so that (see Fig. 4.3.1) P− = P+ − t(L x P− + L y P− ) + t Q f + O(t 2 ),
(4.3.48)
where P− = P(x, y, tk+1 −) and P+ = P(x, y, tk +). Recall that both P+ and P− are independent of k at equilibrium, hence (4.3.40) and (4.3.45) yield P+ = P− − PT− (P− + C )−1 P− .
(4.3.49)
Combining (4.3.48) and (4.3.49) we have t(L x P− + L y P− ) + PT− (P− + C )−1 P− = t Q f + O(t 2 ).
(4.3.50)
Notice the nonlinearity of the impact of data sites upon P. It is possible for P to strike a balance between the two terms on the lefthand side of (4.3.50) (dynamics and dataimpact). This balance can take the form of a boundary layer around data sites. The Kalman gain K and the Kalman ﬁlter estimate w will have this structure, which is quite unphysical (Bennett, 1992). It arises from the adoption of a “cycling” algorithm, as in (4.3.45).
Exercise 4.3.3 Show that there is no such nonlinearity in the nonsequential representer algorithm, for one ﬁxed smoothing interval [0, T ] that may include many measurement times: 0 < t1 < · · · < t n < · · · < t N < T .
Exercise 4.3.4 Consider smoothing a sequence of such intervals: K T < t < (K + 1)T , K = 0, 1, 2, . . . , using the inverse estimate at the end of the K th interval as the ﬁrstguess initial ﬁeld at the start of the (K + 1)th : I K +1 (x) = uˆ K (x, (K + 1)T ), and using the error covariance for the inverse estimate as the error covariance for the ﬁrstguess initial ﬁeld: CiK +1 (x, y) = CuˆK (x, y, (K + 1)T ). Show that the equilibrium error covariance for this “cycling” inverse obeys a nonlinear equation like (4.3.50). Hint: for simplicity,
104
4. The varieties of linear and nonlinear estimation
assume that the domain is inﬁnite: −∞ < x < ∞, assume that the ﬁrstguess forcing ﬁeld F K is perfect: C Kf = 0, and integrate (3.2.43) as crudely as (4.3.48). 4.3.6
“Colored noise”: the augmented Kalman ﬁlter
We may relax the assumption (4.3.5) of “white system noise”. The simplest “colored system noise” has covariance E( f (x, t) f (y, s)) = Q f (x, y)e−
t−s τ
(4.3.51)
for some decorrelation time scale τ > 0. Note that the Q f s appearing in (4.3.5) and (4.3.51) have different units of measurement. It may be shown that (4.3.51) is satisﬁed by solutions of the ordinary differential equation df (x, t) − τ −1 f (x, t) = q(x, t), dt
(4.3.52)
E(q(x, t)q(y, s)) = (τ/2)−1 Q f (x, y)δ(t − s),
(4.3.53)
E( f (x, 0) f (y, 0)) = Q f (x, y),
(4.3.54)
E( f (x, 0)q(y, s)) = 0.
(4.3.55)
provided
and
This suggests augmenting the state variable (Gelb, 1974): u(x, t) u(x, t) → . f (x, t)
(4.3.56)
The dynamical model is now (4.3.46), (4.3.52). Note that the “colored” random process f (x, t) is now part of the state to be estimated. The augmented system is driven by the “white noise” q(x, t). The augmented error covariance now includes crosscovariances of errors in the Kalman ﬁlter estimates of u and f .
4.3.7
Economies
The Kalman ﬁlter is a very popular data assimilation technique, owing to its being sequential (e.g., Fukumori and MalanotteRizzoli, 1995; Fu and Fukumori, 1996; Chan et al., 1996). Also, the “analysis” step (4.3.39) is identical to synoptic or spatial optimal interpolation, as widely practiced already in meteorology and oceanography (Miller, 1996; MalanotteRizzoli et al., 1996; Hoang et al., 1997a; Cohn, 1997). The Kalman ﬁlter algorithm evolves the error covariance P in time, via (4.3.41), and (4.3.45). Nevertheless, evolving P is a massive task for realistically large systems so many compromises are made. For example, the covariance P(x, y, t) is evolved on a computational grid much coarser than the one used for the state estimate w(x, t), or P(x, y, t)
4.4 Maximum likelihood
105
is integrated to an equilibrium covariance P∞ (x, y) which is then used at all times t1 , . . . , t K (Fukumori and MalanotteRizzoli, 1995), or the number of degrees of freedom in w(x, t) is reduced by an expansion in spatial modes (Hoang et al., 1997b). A covariance such as P may also be approximated by statistical simulation, as discussed in §3.2.
4.4
Maximum likelihood, Bayesian estimation, importance sampling and simulated annealing
4.4.1
NonGaussian variability
Leastsquares is the simplest of all estimators. It has so many merits. Gradients and Euler–Lagrange equations are available, so long as the dynamics are smooth. Structural analyses in terms of null spaces, data spaces, representers and sweep algorithms are available, as are statistical closures such as the Kalman ﬁlter, when the dynamics are linear or linearizable. Why, then, choose other estimators? Consider ocean temperatures near the Gulf Stream front. As the latter meanders back and forth across the mooring, the temperature switches rapidly between the higher value for the warm Sargasso Sea water and the lower value for the cool slope water. Thus the frequency distribution of temperature would be bimodal, with peaks at the two values. A leastsquares analysis of temperature would yield the average temperature, which is in fact realized only brieﬂy while the front is passing through the mooring. What would be a more suitable estimator? Can samples of the nonnormal population be generated? How can its estimator be minimized? 4.4.2
Maximum likelihood
Let us review some introductory statistics. Suppose the continuous random variable u has the probability distribution function p(u; θ), where θ is some parameter. Let u 1 , . . . , u n be independent samples of u. Then the joint pdf of the samples is the likelihood function: L(θ ) = p(u 1 , . . . , u n ; θ ) =
n )
p(u i ; θ).
(4.4.1)
i=1
That is, L
n *
du i is the probability that the n samples are in the respective intervals
i=1 du i ),
(u i , u i + 1 ≤ i ≤ n. The maximum likelihood estimate of θ is that value of θ for which L(θ) assumes its maximum value. As an illustration, suppose that u is normally distributed with mean µ and variance σ 2 : p(u; µ, σ ) = (2πσ 2 )− 2 exp[−(2σ 2 )−1 (u − µ)2 ]. 1
(4.4.2)
106
4. The varieties of linear and nonlinear estimation
Note that there are two parameters here: µ and σ . Given n samples of u, what are the maximum likelihood estimates of µ and σ 2 ? The likelihood function is L(µ, σ ) ≡
n )
p(u i ; µ, σ )
i=1
(4.4.3) %
2 −n/2
= (2πσ )
2 −1
exp −(2σ )
n
& (u i − µ)
2
.
(4.4.4)
i=1
We may as well seek the maximum of n l = log L = − log(2πσ 2 ) − (2σ 2 )−1 (u i − µ)2 . 2 i=1 n
(4.4.5)
Extremal conditions are ∂l = −2(2σ 2 )−1 (u i − µ) = 0, ∂µ i=1 n
(4.4.6)
∂l 1 n (u i − µ)2 = 0. = − σ −2 + σ −4 2 ∂(σ ) 2 2 i=1 n
(4.4.7)
The ﬁrst condition yields µ L = n −1
n
ui ,
(4.4.8)
(u i − µ L )2 .
(4.4.9)
i=1
the second yields σ L2 = n −1
n i=1
So µ L , the maximum likelihood estimate for µ, is just the arithmetic mean, while σ L2 is just the sample variance. Now suppose that the pdf for u is exponential, centered at µ and with scale σ : p(u; µ, σ ) = (2σ )−1 exp[−σ −1  u − µ ]. Then
% L(µ, σ ) = (2σ )−n exp −σ −1 l(µ, σ ) = −n log(2σ ) − σ −1
n i=1 n
(4.4.10) &
u i − µ ,
(4.4.11)
u i − µ.
(4.4.12)
i=1
Hence ∂l = −σ −1 ∂µ
µ > ui
1−
µ < ui
1
= 0,
(4.4.13)
4.4 Maximum likelihood
107
provided u i = µ for any i, while ∂l n u i − µ = 0. = − + σ −2 ∂σ σ i=1 n
(4.4.14)
So n
σ L = n −1
u i − µ L ,
(4.4.15)
i=1
but µ L is not so easily determined. If n is even, then µ L should be greater than n/2 samples and less than n/2 + 1. Let’s assume that the samples are ordered: u 1 ≤ u 2 ≤ · · · ≤ u n2 < u n2 +1 ≤ · · · ≤ u n .
(4.4.16)
Then we should choose µ L such that u n2 < µ L < u n2 +1 ;
(4.4.17)
hence −
n
n
u i − µ L  =
i=1
2
(u i − µ L ) −
(u i − µ L )
(4.4.18)
i= n2 +1
i=1 n
=
n
2
ui −
n
ui ,
(4.4.19)
i= n2 +1
i=1
which is independent of µ L ! Had we chosen, say u n2 −1 < µ L < u n2 ,
(4.4.20)
then −
n
2 −1 n
u i − µ L  =
i=1
(u i − µ L ) −
n
=
ui −
n
=
i=1
n
u i + 2µ L
i= n2
i=1 2
(u i − µ L )
i= n2
i=1 2 −1
n
ui −
n
u i − 2(u n2 − µ L ).
(4.4.21)
i= n2 +1
But the rhs of (4.4.21) is less than the rhs of (4.4.19) by 2(u n/2 − µ L ), which is positive by virtue of (4.4.20). So the choice (4.4.17) is maximal. Note that µ L is only determined within the interval u n/2 < µ L < u n/2+1 , and that σ L is insensitive to the choice. Maximum likelihood estimation is trivial for normal distributions, but less so for others. Returning to the normal case, notice that we chose µ to maximize l, as in (4.4.5), that is, to minimize N i=1
(u i − µ)2
(4.4.22)
108
4. The varieties of linear and nonlinear estimation
with respect to µ. The maximum likelihood estimate of µ is the leastsquares estimate. Now let’s change the perspective slightly. Suppose that u 1 , . . . , u n are n measurements of a quantity u, and each measurement involves an error i ≡ u i − u that is independent of the other errors, and distributed as p(; ν, θ) = (2πθ 2 )−1/2 exp[−(2θ 2 )−1 ( − ν)2 ],
(4.4.23)
where the mean ν and variance θ 2 are known. That is, p(u i ; ν + u, θ ) = (2πθ 2 )−1/2 exp[−(2θ 2 )−1 (u i − u − ν)2 ].
(4.4.24)
Exercise 4.4.1 Show that u L , the maximum likelihood estimate of u, is u L = n −1
n
(u i − ν).
(4.4.25)
i=1
Remove the instrument bias, and take the arithmetic mean! Note that Eu L = n −1
n
Eu i − ν
i=1
= n −1
n
Ei + u − ν
i=1
= ν+u−ν = u,
(4.4.26)
where ∞ Ei ≡
p(i , ν, θ)i di .
(4.4.27)
−∞
The result (4.4.26) tells us that u L is an unbiased estimate of u.
Exercise 4.4.2 Show that E((u L − u)2 ) = θ 2 .
(4.4.28)
That is, the variance of the error in the maximum likelihood estimate u L for u is equal to the variance of the measurement errors. Now let’s consider randomly erroneous measurements of a random quantity: v = u + ,
(4.4.29)
4.4 Maximum likelihood
109
where u is a random variable and a random measurement error. We shall denote their pdfs as p(u) and p(). (This is a poor notation; pu (x) and p (x) would be better, where pu (x)d x is the probability that x < u < x + d x and p (x)d x is the probability that x < < x + d x, but let’s try to keep notation simple if imprecise.) The joint pdf for both u AND is p(u, ). Then the marginal pdfs are p(u) = p(u, ) d, p() = p(u, ) du. (4.4.30) We may also consider the conditional distributions p(u) and p(u). The former is the probability distribution of u, given a value for ; the latter is the probability distribution of , given a value of u. Hence, p(u, ) = p(u) p() = p(u) p(u).
(4.4.31)
If u and are independent, then p(u) = p(u),
p(u) = p(),
(4.4.32)
and (4.4.31) reduces to the product rule: p(u, ) = p(u) p().
(4.4.33)
In the general case, where u and may be dependent, (4.4.31) becomes Bayes’ Rule (Cox and Hinkley, 1974): p(u) = p(u)
p(u) . p()
Combining (4.4.30) and (4.4.31) yields p(u) = p(u) p() d,
p() =
Combining (4.4.34) and (4.4.35) yields p(u) = p(u) p(u) d,
p() =
that is,
(4.4.34)
p(u) p(u) du.
(4.4.35)
p(u) p() du,
(4.4.36)
p(u) d =
p(u) du = 1.
(4.4.37)
Exercise 4.4.3 Examine “Egbert’s Table” (see next page) of values for p(u, ) for a simple case in which u = 1 or 2, while = −1, 0 or 1. The upper number in each of the six boxes is p(u, ).
110
4. The varieties of linear and nonlinear estimation
1 u 2
−1
0
1
0.09
0.72
0.09
0.1 0.05
0.0
0.05
p(u)
p()
(i) Calculate p(u) for each u, and p() for each . Verify that these distributions are normalized. (ii) Use (4.4.31) to calculate p(u). Enter the values in the six boxes (a check value is provided in the ﬁrst box). (iii) Show that var ( = 0.28), var (u = 2) = 1.0. That is, an “observation” of the random variable u may increase the variance of an unknown but dependent random variable ! Now suppose that we have n independent data v1 , . . . , vn . We wish to form the n * likelihood function L = p(vi ). We can determine p(v), if we know p(vu) and i=1
p(u). But p(vu) is the pdf for recording the value v when the true value is u. That is, p(vu) is the pdf for , when = v − u. At this point our sloppy notation fails us, and we must write p(vu) = p (v − u). Then
(4.4.38)
p(v) =
p (v − u) p(u) du.
(4.4.39)
The distributions in the integrand have parameters E, σ2 , Eu, σu2 , . . . which we would like to estimate, using the data v1 , . . . , vn . We may do so, by solving the maximum likelihood conditions ∂ ln L = 0, etc. (4.4.40) ∂ E Thus we arrive at maximum likelihood estimators, given conditional and marginal distributions.
Exercise 4.4.4 Assume that p and pu are normal. Derive the maximum likelihood estimates of the four parameters, given independent data v1 , . . . , vn , where v = u + .
4.4 Maximum likelihood
4.4.3
111
Bayesian estimation
Let us now apply these ideas to optimal interpolation (Lorenc, 1997). The gridded multivariate ﬁeld can be ordered as a vector of N components: u = (u 1 , . . . , u N ) ∈ R N . Assume that we have a prior or “background” estimate ub , which is usually a model solution. The prior conditional distribution is p(uub ). Let there again be M measurements (v1 , . . . , v M ) ∈ R M related to the true ﬁeld by v = H(u) + , where the independent error is = (1 , . . . , M ) ∈ R M and has the pdf p (; E, C , . . .). Thus H maps R N into R M . If it is linear, then we may write H(u) = Hu, where H is an M × N matrix. We want the distribution for u, given the background ub and data v. Bayes’ Rule becomes p(uv, ub ) =
p(vu, ub ) p(uub ) . p(vub )
(4.4.41)
Notice that ub is being regarded as a ﬁxed parameter here; only u and v are being interchanged. Moreover, the measurement process is unrelated to the model, so p(uv, ub ) =
p(vu) p(uub ) p(v)
(4.4.42)
∝ p (v − H(u); E, C , . . .) p(uub ).
(4.4.43)
We may ignore the denominator, as it is independent of u. Given (4.4.43), we take as our “analysis” estimate of u the mean value: ! u p(uv, ub ) du u = ! . p(uv, ub ) du a
(4.4.44)
Exercise 4.4.5 Suppose that both distributions in (4.4.44) are normal; that is, p (; . . .) = N (; E, C ),
(4.4.45)
p(uu ) = N (u; u , Cuu ).
(4.4.46)
b
b
Derive the standard leastsquares optimal interpolation formulae from (4.4.44).
In summary, if we can choose credible distributions for the data error and the background error, be they normal or otherwise, we can use Bayes’ Rule to construct a pdf for the ﬁeld. Its ﬁrst moment is a credible estimate of the ﬁeld.
Exercise 4.4.6 Or is it?
4. The varieties of linear and nonlinear estimation
112
4.4.4
Importance sampling
Not all oceanic and atmospheric processes are normally distributed. Not all dynamics and penalty functionals are smooth. There is a need for estimators other than leastsquares, and for optimization methods other than the calculus of variations. But ﬁrst we need a method for synthesizing samples from any probability distribution P. More precisely, we require an algorithm for generating a sequence of real numbers x1 , x2 , . . . , xn , . . . such that the number of values in the interval (x, x + d x) is proportional to P(x) d x. That is, we wish to perform “importance sampling”. It is usually the case that P is in fact a normalized probability distribution function, that is, P(x) = K −1 Q(x),
(4.4.47)
where b K =
Q(x) d x
(4.4.48)
a
for some interval a ≤ x ≤ b. We often only know Q(x) and would like to avoid evaluation of K , especially for higher dimensional problems in which (4.4.48) is a multiple integral. Consider a Markov chain x1 , x2 , . . . , xn , . . . , for which the value of xn+1 lies in the interval a ≤ x ≤ b, depends only upon the value of xn and the dependence is random. Let Pn (x)d x be the probability that x < xn < x + d x, and let T (x, y)d x d y be the probability that x < xn+1 < x + d x given that y < xn < y + dy. Thus, T is a transition probability density, and y=b {T (x, y)Pn (y) − T (y, x)Pn (x)} d y d x. Pn+1 (x) d x = Pn (x) d x +
(4.4.49)
y=a
The ﬁrst term in the integrand accounts for transitions to x from all possible y; the second accounts for transitions from x to all possible y. Note that the integral is over y. The chain is in equilibrium if Pn+1 (x) = Pn (x), which implies that the integral in (4.4.49) vanishes. That is the condition of balance. Note the assumption that T is independent of n. The condition of detailed balance: T (x, y)P(y) − T (y, x)P(x) = 0
(4.4.50)
is sufﬁcient but not necessary for equilibrium. Then the Markov chain x1 , x2 , . . . is in equilibrium, with distribution P(x). A simple algorithm for generating a chain from a given pdf P is as follows (Metropolis et al., 1953). (1) Pick a number z at random in [a,b]. (2) Calculate r=
P(z) . P(xn )
(4.4.51)
4.4 Maximum likelihood
(3) Pick a number η at random in [0,1]. (4) Choose z, xn+1 = xn , In effect the choice is
xn+1 =
if
η < r;
if
η > r.
z,
with probability r ;
xn ,
with probability 1 − r.
113
(4.4.52)
(4.4.53)
Hence if r < 1 then r is the probability of transition from xn to z: T (z, xn ) = r =
P(z) , P(xn )
(4.4.54)
while if r > 1, then T (z, xn ) = 1.
(4.4.55)
The condition of detailed balance follows immediately. For example, if r < 1: T (z, xn )P(xn ) =
P(z) P(xn ) = P(z) = T (xn , z)P(z). P(xn )
(4.4.56)
Note that the algorithm depends on P only via the ratio (4.4.51). In fact r=
Q(z) . Q(xn )
(4.4.57)
It is not necessary to know the normalizing constant (4.4.48).
4.4.5
Substituting algorithms
Suppose that u is the solution of an equation such as D(u) = f,
(4.4.58)
where D is some nonlinear function, while f is a random variable with pdf A = A( f ). That is, the probability of g < f < g + dg is A(g)dg. What is the corresponding pdf B = B(u)? This is a nontrivial analytical problem if D is nontrivial. However, we can construct a Markov chain u 1 , u 2 , . . . distributed according to B. Use importance sampling, based on the nonnormalized pdf Q(u) = A(D(u)). For a given u n , pick z at random, calculate r = Q(z)/Q(u n ), and proceed as in (4.4.52). We would then have properly distributed samples for u, and could form a histogram estimate of its pdf B. These samples are now being loosely described as “ensembles” in the literature. The ensemble is the totality. Note especially that we do not have to invert the operator D. We merely have to evaluate it for each u n , by direct substitution of u n into D.
114
4. The varieties of linear and nonlinear estimation
Exercise 4.4.7 Estimate the pdf of u, given that loge u = f
(4.4.59)
and that f is Gaussian.
This approach may be invaluable when f is a random ﬁeld and D represents the dynamics of an ocean model. Steady models can be particularly difﬁcult to solve, especially if they are nonlinear. Some timedependent intermediate models include diagnostic equations that are unwieldy. An obvious example is the stratiﬁed quasigeostrophic model. In particular, diagnosing (solving) the threedimensional elliptic equation ∇ 2 ψ = ξ for the streamfunction ψ in a realistic ocean basin is nontrivial: assembling sparse matrices requires great care. In comparison, it is relatively trivial to substitute the streamfunction into the elliptic equation, and then substitute the vorticity ξ into the ﬁrstorder wave equation: ∂ξ + J (ψ, ξ + βy) = q, ∂t
(4.4.60)
where β is the local meridional gradient of the Coriolis parameter, and where q is some random source of vorticity.
4.4.6
Multivariate importance sampling
Thus far, u has been a single, real random variable. We are interested in random multivariate ﬁelds: u = u(x, y, z, t), v = v(· · ·), w = w(· · ·), p = p(· · ·), etc. In computational practice, these ﬁelds are deﬁned on grids, thus we have arrays u i jkl = u(xi , y j , z k , tl ), vi jkl = v(· · ·), wi jkl = w(· · ·), pi jkl = p(· · ·), etc. For clarity, let us condense all these into a single vector u = (u 1 , . . . , u m , . . . , u M ). A Markov chain of these vectors will be denoted by un = (u n1 , u n2 , . . . , u nM ), for n = 1, 2, 3, . . . . Notice that the upper index n is not the time index; the latter is included in the lower index. Suppose that the multivariate probability distribution for u is factorable: Q(u) = Q 1 (u 1 )Q 2 (u 2 ) . . . Q M (u M ),
(4.4.61)
in which case the components of u are independent. Consider, for example: ' ( Q(u) = exp − u 21 − u 22 − · · · − u 2M .
(4.4.62)
Then we may apply importance sampling to each component independently. The decision to accept a new value z m for u n+1 would be based on the ratio m rm =
Q m (z m ) . Q m u nm
(4.4.63)
4.4 Maximum likelihood
115
These M decisions could be made in series or in parallel. Now suppose that components of u depend only upon two nearest neighbors: Q(u) = Q 2 (u 1 , u 2 , u 3 )Q 3 (u 2 , u 3 , u 4 )Q 4 (u 3 , u 4 , u 5 ) . . . Q M−1 (u M−2 , u M−1 , u M ). (4.4.64) Consider, for example: Q(u) = exp[−(u 1 + u 3 − 2u 2 )2 − (u 2 + u 4 − 2u 3 )2 − · · · − (u M−2 + u M − 2u M−1 )2 ].
(4.4.65)
Then importance sampling may be performed in three “sweeps”. Assume that M is divisible by three. n+1 n+1 n+1 Sweep 1. Choose trial values z 1 , z 4 , z 7 , . . . , z M−2 for u n+1 1 , u 4 , u 7 , . . . , u M−2 respectively; each decision to accept a trial value is independent of the others. For example, the ratio for sampling u n+1 is 1 Q 2 z 1 , u n2 , u n3 , r1 = (4.4.66) Q 2 u n1 , u n2 , u n3
while for u n+1 it is 4
Q 3 u n2 , u n3 , z 4 Q 4 u n3 , z 4 , u n5 Q 5 z 4 , u n5 , u n6 r4 = Q 3 u n2 , u n3 , u n4 Q 4 u n3 , u n4 , u n5 Q 5 u n4 , u n5 , u n6
(4.4.67)
and so on, for r7 , . . . , r M−2 . All these decisions can be made in parallel. n+1 n+1 Sweep 2. Generate u n+1 2 , u 5 , . . . , u M−1 by importance sampling, in parallel. n+1 n+1 Sweep 3. Generate u 3 , u 6 , . . . , u n+1 M by importance sampling, in parallel. This procedure, known as “checkerboarding”, is complicated when the actual computational grid involves more than one dimension.
4.4.7
Simulated annealing
Consider for simplicity a scalar variable u, for which there is a penalty function J (u). Assume only that J is bounded below: J (u) > B
(4.4.68)
for all u. We wish to ﬁnd the value of u for which J is least. Let u n be an estimate, for which Jn ≡ J (u n ) is unacceptably large. Make a small perturbation to u n : z = u n + u n . The “downhill strategy” is: u n+1 =
z,
if
J (z) < Jn
un ,
if
J (z) > Jn .
(4.4.69)
(4.4.70)
116
4. The varieties of linear and nonlinear estimation
However, this strategy could terminate at a local minimum of J . It would be better to allow a few uphill searches at ﬁrst, in order to avoid such an outcome. So, use importance sampling: if J (z) < Jn z, u n+1 = z, if J (z) > Jn , with probability r u n , if J (z) > Jn , with probability 1 − r, where e−J (z)/θ = e−(J (z)−Jn )/θ < 1 (4.4.71) e−Jn /θ for some positive “annealing temperature” θ . Simply pick a random variable η in [0,1]. If η < r , accept z. If r < η, keep u n . Now r → 0 as θ → 0, so fewer uphill steps are allowed as θ decreases. The “annealing strategy”, or rate of decrease of θ, is a “black art” (Azencott, 1992). Note that no gradient information for J is used. The penalty function need not even be continuous in u. This approach should be ideal for data assimilation with “small” nonlinear biological models that have “switches”. These models typically describe the temporal evolution of a biological system, at one point. The models include constraints such as lower bounds on biomass, together with discontinuous representations of very rapidly adjusting processes. Barth and Wunsch (1990) used simulated annealing to optimize the locations of acoustic transceivers in an idealized model “ocean”. Kruger (1993) used simulated annealing to minimize the penalty functionals associated with the inversion of two ocean models. The ﬁrst was a twobox model of ocean stratiﬁcation that included a nonsmooth representation of convective adjustment. The second involved a singlelevel quasigeostrophic model much like (4.4.60), at one time. There were about 4000 computational degrees of freedom in Kruger’s second application. Importance sampling is used extensively in theoretical physics, especially for the evaluation of path integrals. The number of dimensions is extremely large, so the efﬁcient generation of independent trial values is of crucial importance. Ingenious techniques such as “Hybrid Monte Carlo” or HMC, have been devised, but these typically assume that the integrand depends smoothly upon the state variables. See Chapter 6 for an application of these techniques to the resolution of an illposed problem. r=
Chapter 5 The ocean and the atmosphere
Seawater and air are viscous, conducting, compressible ﬂuids. Yet largescale oceanic and atmospheric circulations have such high Reynolds’ numbers and such low aspect ratios that viscous stresses, heat conduction and nonhydrostatic accelerations may all be neglected. (The Mach number of ocean circulation is so low that the compressibility of seawater may also be neglected, but will be retained here in the interest of generality.) Subject to these approximations, the Navier–Stokes equations simplify to the socalled “Primitive Equations”. It is often convenient to express these equations in a coordinate system that substitutes pressure for height above or below a reference level. The Primitive Equations were for many years too complex for operational forecasting. They were further reduced by assuming low Rossby number ﬂow, leading to a single equation for the propagation of the vertical component of vorticity – the “quasigeostrophic” equation. Now obsolete as a forecasting tool, this relatively simple equation retains great pedagogical value. To its credit, it is still competitive at predicting the tracks of tropical cyclones, if not their intensity. The astronomical force that drives the ocean tides is essentially independent of depth, and so its effects may be modeled by unstratiﬁed Primitive Equations: the Laplace Tidal Equations. The external Froude number for the tides is so low that the “LTEs” are essentially linear. Combining the linear LTEs with the vast records of sea level elevation collected by satellite altimeters makes an ideal ﬁrst test for inverse ocean modeling. The interaction of harmonic analysis of the tides and biasfree strategies for measurement leads to novel measurement functionals. The great separation of scales clariﬁes the prior analysis of errors in the dynamics and in coastlines. Initializing a quasigeostrophic model for hurricane track prediction is ideal as a ﬁrst application of inverse ocean modeling to nonlinear dynamics. Errors arise in the
117
118
5. The ocean and the atmosphere
dynamics owing to the neglect of resolvable processes. The statistics of these processes may be estimated from archived data. The relatively simple “QG” dynamics also clarify discussion of the conceptual issue of even deﬁning statistics for errors arising from parameterization of unresolvable processes. Highaltitude winds inferred from tracks of cloud images collected by satellites cannot reasonably be assimilated into a “QG” model; a Primitive Equation model is called for. This provides an extreme exercise in deriving and solving Euler–Lagrange equations for a variational principle that is based on a complex model and on an unstructured data set. The transPaciﬁc array of instrumented ocean moorings known as TAO is so regularly structured, and of such continuity and duration, that rigorous testing of models becomes feasible. An intermediate model of the seasonaltointerannual variability of the coupled ocean–atmosphere is TAO’s ﬁrst victim. The chapter closes with notes on a selection of contemporary variational and statistical assimilations, variously involving components of the entire hydrosphere.
5.1
The Primitive Equations and the quasigeostrophic equations
5.1.1
Geophysical ﬂuid dynamics is nonlinear
Our development of inverse theory has involved linear models and linear measurement functionals. Tides provide a splendid example of a linear model, but there are no others. In general, geophysical ﬂuid dynamics is nonlinear. The Primitive Equations and the quasigeostrophic equations of motion (Haltiner and Williams, 1980; Gill, 1982) will be brieﬂy reviewed in this section.
5.1.2
Isobaric coordinates
Let us replace space–time coordinates (X, Y, Z , T ) with space–pressure–time coordinates (x, y, p, t), where x = X,
y = Y,
p = p(X, Y, Z , T )
and
t = T.
(5.1.1)
Note 1. We could instead be using, say, spherical coordinates (longitude and latitude) on horizontal surfaces (constant Z ), or on isobaric surfaces (constant p) as in (5.1.1). Note 2. The coordinate transformation (5.1.1) depends upon the state of the ocean or the atmosphere, through the instantaneous pressure ﬁeld p.
5.1 Primitive Equations
5.1.3
119
Hydrostatic balance, conservation of mass
In space–time coordinates, the hydrostatic approximation is ∂p = −ρg, ∂Z
(5.1.2)
where ρ is the ﬂuid density, and g the local gravitational acceleration. We may use (5.1.1) and (5.1.2) to obtain the volume element: d x d y d p = ρg d X dY d Z .
(5.1.3)
Now consider a parcel of ﬂuid, occupying a region V = V(t) that moves and distorts in time. The total mass of the parcel does not change, so d ρ d X dY d Z = 0, (5.1.4) dT V
or, as a consequence of (5.1.1) and (5.1.3), d d x d y d p = 0. dt
(5.1.5)
V
Comparing the volume integral at times t and t + dt leads easily to the conclusion that v · nˆ da = 0, (5.1.6) S
where S is the surface of the parcel, nˆ is an outward unit normal on S, da is an element of area in (x, y, p) coordinates, and v = (u, v, ω), where u≡
Dx , Dt
v≡
Dy , Dt
ω≡
Dp . Dt
(5.1.7)
The divergence theorem in (x, y, p) coordinates yields V
∂u ∂v ∂ω + + ∂x ∂y ∂p
d x d y d p = 0.
(5.1.8)
The parcel V is arbitrary, hence the ﬂow is volumeconserving in (x, y, p) coordinates: ∂u ∂v ∂ω + + =0 ∂x ∂y ∂p
(5.1.9)
5. The ocean and the atmosphere
120
5.1.4
Pressure gradients
The pressure gradient per unit mass is ρ −1
∂p = −g ∂X
∂p ∂Z
−1
∂p , ∂X
(5.1.10)
by virtue of the hydrostatic approximation (5.1.2). The chain rule applied to Z = Z (x, y, p, t) yields 1=
∂Z ∂ Z ∂x ∂ Z ∂y ∂ Z ∂p ∂ Z ∂t = + + + . ∂Z ∂x ∂ Z ∂y ∂ Z ∂p ∂ Z ∂t ∂ Z
(5.1.11)
But x = X , y = Y and t = T are orthogonal to Z , hence ∂x ∂y ∂t = = = 0, ∂Z ∂Z ∂Z
(5.1.12)
and so
∂p ∂Z
−1
=
∂Z . ∂p
(5.1.13)
Note that (5.1.13) is not a general property of transformations; it is only true for our special transformation (5.1.1). The hydrostatic approximation then becomes ∂φ = −ρ −1 ∂p
(5.1.14)
where we have deﬁned the geopotential φ = φ(x, y, p, t): φ ≡ gZ.
(5.1.15)
Combining (5.1.10) and (5.1.13) yields ρ −1
∂p ∂ Z ∂p = −g . ∂X ∂p ∂ X
(5.1.16)
Next we use the chain rule on Z = Z (x, y, p, t) to obtain 0= But
∂x ∂X
= 1,
∂y ∂X
∂Z ∂ Z ∂x ∂ Z ∂y ∂ Z ∂p ∂ Z ∂t = + + + . ∂X ∂x ∂ X ∂y ∂ X ∂p ∂ X ∂t ∂ X
= 0 and
∂t ∂X
(5.1.17)
= 0, so (5.1.10) becomes ρ −1
∂p ∂Z ∂φ =g = . ∂X ∂x ∂x
(5.1.18)
5.1 Primitive Equations
121
Similarly, ρ −1
∂p ∂φ = . ∂Y ∂y
(5.1.19)
Exercise 5.1.1
Draw a sketch that explains (5.1.18) and (5.1.19).
5.1.5
Conservation of momentum
Now x = X and t = T , so DX D2x Du Dx DU D2X = = = ≡ u, = , DT Dt DT DT 2 Dt 2 Dt where u = (u, v). Thus the horizontal momentum equation U≡
DU ˆ × U = −ρ −1 ∇ X p, ∼ + f k DT
(5.1.20)
(5.1.21)
∂ where ∇ X = ∂∂X , ∂Y and f = f (Y ) is the (known) Coriolis parameter, becomes the isobaric momentum equation Du + f kˆ × u = −∇x φ Dt
(5.1.22)
where
and ∇x = 5.1.6
∂ , ∂ ∂x ∂y
D ∂ ∂ = + u · ∇x + ω , Dt ∂t ∂p
(5.1.23)
.
Conservation of scalars
For any conserved tracer τ such as entropy η, salinity S or relative humidity q, Dτ = 0. DT But T = t, so Dτ =0 Dt
(5.1.24)
We have now derived the Primitive Equation in pressure coordinates: (5.1.9), (5.1.14), (5.1.22) and (5.1.24). Note that only the hydrostatic approximation has been made;
122
5. The ocean and the atmosphere
incompressibility has not been assumed. The appearance of (5.1.9) is deceptive! The Primitive Equations must be supplemented with an equation of state such as ρ = ρ( p, η, S)
(5.1.25)
ρ = ρ( p, η, q)
(5.1.26)
or
where the rhs of (5.1.25), (5.1.26) indicates a prescribed functional form.
Exercise 5.1.2 Consider hydrostatic motion of a ﬂuid of constant density, between a rigid ﬂat surface at φ = 0 and a material free surface at φ = gh, where h = h(x, y, t). Assume that the pressure at the free surface vanishes: p = 0. Derive the “shallow water equations”: Du + f kˆ × u = −g∇h, Dt Dh + h∇ · u = 0. Dt Note: It is dynamically consistent to assume that ∂u/∂ p = 0. 5.1.7
(5.1.27) (5.1.28)
Quasigeostrophy
The Eulerian form of (5.1.22) is ∂u ∂u + (u · ∇)u + ω + f kˆ × u = −∇φ, ∂t ∂p
(5.1.29)
where it is now understood that ∇ is ∇x . For mesoscale motions and larger, it is reasonable to assume that ∂ω ∇ · u , ω ∂u u · ∇u , (5.1.30) ∂p ∂p hence the Primitive Equations (5.1.9) and (5.1.22) are approximately ∇ · u = 0,
(5.1.31)
∂u + (u · ∇)u + f kˆ × u = −∇φ, (5.1.32) ∂t which is the same as the dynamics of planar incompressible ﬂow. Note that ∇ is a gradient at constant pressure, and that (5.1.31) and (5.1.32) form a closed system, to the extent that they determine u and φ without reference to the equation of state, or to the hydrostatic approximation or to the conservation of entropy. As is well
5.1 Primitive Equations
123
known, in simplyconnected domains (5.1.31) implies the existence of a streamfunction ψ = ψ(x, p, t) such that ∂ψ ∂ψ ˆ u = k × ∇ψ = − , . (5.1.33) ∂y ∂x If we deﬁne the vertical component of relative vorticity by ξ ≡ kˆ · ∇ × u =
∂v ∂u − = ∇ 2 ψ, ∂x ∂y
(5.1.34)
and if we apply kˆ · ∇× to (5.1.32) and use (5.1.31), we obtain the vorticity equation ∂ξ df + u · ∇ξ + v = 0, ∂t dy
(5.1.35)
∂ζ + u · ∇ζ = 0 ∂t
(5.1.36)
ζ =ξ+ f
(5.1.37)
or
where
is the total vorticity. The conservation law (5.1.36) may be expressed entirely in terms of the streamfunction: ∂ 2 ∂(ψ, ∇ 2 ψ + f ) ∇ ψ+ =0 ∂t ∂(x, y)
(5.1.38)
The nonlinearity of this “ﬁltered” vorticity equation is signiﬁcant for large Rossby number: Roβ ≡
U 1, βl 2
(5.1.39)
where β ≡ ddyf , while U and l are the scales of variation of u and x respectively. Note that (5.1.39) may still hold even though Ro f ≡
U
1, f 0l
βl f 0 ,
(5.1.40)
where f 0 is a local value of f . Under these conditions, (5.1.29) yields the “geostrophic” balance: f 0 kˆ × u ∼ = −∇φ,
(5.1.41)
124
5. The ocean and the atmosphere
in which case (5.1.31) again holds approximately, and (5.1.35), (5.1.36) may be derived from (5.1.29) at O(Ro f ). As a consequence, (5.1.36) is also known as the “quasigeostrophic” vorticity equation (Gill, 1982).
5.2
Ocean tides
5.2.1
Altimetry
Barotropic ocean tides are globalscale motions that are accurately modeled with linear dynamics. The TOPEX/POSEIDON altimetric satellite, launched in August 1992 and still in operation (June, 2001), is providing highly accurate global sealevel data (Chelton et al., 2001). There could hardly be a more elegant exercise in data assimilation. Indeed, altimetry provided a ﬁrst test of an advanced method using a large amount of data. Let us pause in our development of inverse methods, and explore the real problem of global ocean tides.
5.2.2
Lunar tides
Let us brieﬂy review lunar tides. They are driven by the gravitational attraction of the moon: see Fig. 5.2.1. At the earth’s center of mass E, there is exact equality between the gravitational attraction towards the center of mass of the moon at C and the centripetal acceleration of E towards C, as E moves tangentially on its orbit around C. (More precisely, C and E orbit around the common center of mass.) Let the points A and B make orbits of the same radius as that of E, and so have the same centripetal acceleration as E. However, A is closer to C than is E (while B is further away), and so A experiences a stronger gravitational acceleration towards C (while B experiences a
Figure 5.2.1 Tidal potentials.
N M C A
E
B L
5.2 Ocean tides
125
weaker gravitational acceleration). Hence there are net accelerations or “tidal bulges” at A and B, respectively towards and away from C. Meanwhile the earth spins around its polar axis once a day, by deﬁnition of the polar axis and the day. So each point on the ocean surface should have two high tides (A and B) and two low tides (L and M ) each day. In fact, the ocean has free barotropic motions with many periods of the order of a day, hence its response to the tidegenerating force is very complicated. Solar tides add to the complexity. For a marvelous account of tides in the ocean, see Cartwright (1999). The tidegenerating force (tgf ) is conservative, and so there is a tidegenerating potential (tgp) per unit mass which we shall express as a seasurface elevation h. The tgf is ∇h. In midocean, h is about 30 cm. The tgf has a complicated time dependence, dominated by the relative motions of the earth, the moon and the sun. Certain periodicities are obvious, and up to 400 others have been calculated using celestial mechanics, by G. Darwin, Doodson and others. That is, h(x, t) ∼ = Re
K
h k (x)eiωk t ,
(5.2.1)
k=1
where x denotes a position on the earth’s surface. The frequencies ω1 , . . . , ω K deﬁne tidal constituents. For example ω1 ≡ ‘M2 ’, the “principal lunar semidiurnal constituent”, corresponds to a period of 12h 25m 42s approx., while ω2 ≡ ‘S2 ’, the “principal solar semidiurnal constituent”, corresponds to a period of 12h exactly. Table 5.2.1 is extracted from Doodson and Warburg (1941). The “speed number” is the frequency ωk expressed in degrees per hour (and is equal to exactly 30 for S2 ), while the “relative coefﬁcient” is the relative amplitude of h k . The dominant diurnal and semidiurnal constituents are, in order, M2 , K 1 , S2 , O1 , P1 , N2 , K 2 and Q 1 . The lunar fortnightly, monthly and solar semiannual tides M f , Mm and Ssa are also signiﬁcant. Constructive and destructive interference between semidiurnal and diurnal tides causes a diurnal inequality, that is, one of the two daily high tides exceeds the other. Interference between semidiurnal tides, especially between M2 and S2 , causes beating or “neap” and “spring” tides.
Exercise 5.2.1 What is the period of beats between M2 and S2 ?
5.2.3
Laplace Tidal Equations
Having brieﬂy reviewed the tidegenerating force, let us now review ocean hydrodynamics. It sufﬁces to consider the linear, shallowwater equations on a rotating planet
126
5. The ocean and the atmosphere
Table 5.2.1 List of harmonic constituents of the equilibrium tide on the Greenwich Meridian
Symbol
Argument
Sa Ssa Mm MSf Mf
h 2h s−p 2s − 2h 2s
K1 O1 P1 Q1 ˜1 M
15◦ t 15◦ t 15◦ t 15◦ t 15◦ t 15◦ t 30◦ t 30◦ t 30◦ t 30◦ t 30◦ t 30◦ t 30◦ t 30◦ t 30◦ t
J1 M2 S2 N2 K2 ν2 µ2 L2 T2 2N2
+ h + 90◦ + h − 2s − 90◦ − h − 90◦ + h − 3s + p − 90◦ + h − s + 90◦ + h + s − p + 90◦ + 2h − 2s + 2h − 3s + p + 2h + 4h − 3s − p + 4h − 4s + 2h − s − p + 180◦ − h + p + 2h − 4s + 2 p
Speed number
Relative coefﬁcient
0.0411 0.0821 0.5444 1.0159 1.0980
0.012 0.073 0.083 0.014 0.156
15.0411 13.9430 14.9589 13.3987 14.4921 15.5854 28.9841 30.0000 28.4397 30.0821 28.5126 27.9682 29.5285 29.9589 27.8954
0.531 0.377 0.176 0.072 0.040 0.030 0.908 0.423 0.174 0.115 0.033 0.028 0.026 0.025 0.023
(the Laplace Tidal Equations or LTEs). In the f plane approximation (Gill, 1982), these are ∂u + f kˆ × u = −g∇(h − h) − r u/H, ∂t ∂h + ∇ · (H u) = 0, ∂t
(5.2.2) (5.2.3)
where f is the local value of the Coriolis parameter, kˆ is the unit vector in the local verticallyupward direction, u = u(x, t) is the barotropic current, h = h(x, t) is the sealevel disturbance, H = H (x) is the mean depth of the ocean, r is a bottom drag coefﬁcient and h = h(x, t) is the tgp: see Fig. 5.2.2. Note 1. A quadratic drag law −kuu is more reliable. Note 2. If h ≡ h, then the ocean is in hydrostatic balance with the tgf: this is the “equilibrium” tide of Newton. For longperiod tides such as M f , Mm and Ssa, it is an excellent approximation.
5.2 Ocean tides
127
Figure 5.2.2 Shallowwater theory.
h h
H
Note 3. Many more effects can be included, yielding real gains in forecast accuracy. These include (i) load tide: as sea level rises and falls, the ocean ﬂoor subsides and rebounds elastically; (ii) earth tides: the tgf directly drives motions in the elastic earth; (iii) self tide: as sea levels rises, the local accumulation of mass deﬂects the local vertical; (iv) geoid corrections: the earth tides change the shape of the earth and hence that of the earth’s geopotentials or “horizontals”; (v) atmospheric tides: the tgf and solar heating drive motions in the atmosphere which perturb sealevel pressure. The LTEs require boundary conditions, such as u · nˆ = 0
(5.2.4)
h = hB
(5.2.5)
at coasts, or
at an open boundary. These are unsatisfactory: is the boundary at the shore line or the shelf break? Can h B be measured economically? How shall we avoid spurious oscillations in an open region, when the LTEs are subjected to (5.2.5)? 5.2.4
Tidal data
Tides are the best measured of all ocean phenomena. The data include: (i) centurylong highquality time series of sea level at about one hundred coastal stations, measured with ﬂoats in “stilling wells” and stripchart recorders; (ii) yearlong highquality time series of bottom pressure in about twenty deep ocean locations, measured with the piezoelectric effect and digital recorders; (iii) yearlong goodquality time series of ocean current at selected depths at about a thousand deep locations;
128
5. The ocean and the atmosphere
Figure 5.2.3 Reciprocalshooting tomography.
X(s2)
X(s)
s2
d = ∫ u (X( s), t ) ⋅ dX ( s) + ∈ s1
X(s1)
orbit
Figure 5.2.4 Satellite radar altimetry.
GHz radar traveltime sea level
(iv) yearlong goodquality time series of reciprocalshooting acoustic tomography at about a dozen deep ocean locations: see Fig. 5.2.3. (v) satellite altimetry: see Fig. 5.2.4. Altimetric missions include GEOSAT, ERS1 and TOPEX/POSEIDON. The last (T/P) has a tenday repeattrack orbit from 70◦ S to 70◦ N approx.: see Fig. 5.2.5. Again, TOPEX/POSEIDON has been ﬂying and operating successfully since August 1992. Temporal variability in the orbit of T/P is known with remarkable precision: ±2 cm. However, the shape of the gravitational equipotential (the “geoid”) is not known so accurately. This bias can be eliminated from the data by considering “crossover” differences: see Fig. 5.2.6. The datum becomes d = h(X, TD ) − h(X, T A ) + . Note that T A and TD need not be the times of consecutive passes over X. The values of T A − TD  for consecutive passes can be as large as ﬁve days, thus semidiurnal tides are severely aliased. In fact, the aliased tides resemble Rossby waves with periods of about 60 days. We shall use the dynamics of the LTE to identify and hence reject the aliased tides, which have great spatial coherence.
Figure 5.2.5 TOPEX orbits.
130
5. The ocean and the atmosphere
Ascending (A)
Figure 5.2.6 Orbit crossovers.
N
X
Descending (D)
5.2.5
The state vector, the crossover measurement functional and the penalty functional
Let us work with tidal volume transports Uk (x) ≡ H (x)uk (x) and elevations h k (x) at frequency ωk , for 1 ≤ k ≤ K . Then the state vector ﬁeld is U1 h1 U2 h2 3K U = U(x) = (5.2.6) ... ∈ C . . .. UK hK Recall that Uk and h k are complexvalued. The dynamics become iωk Uk + f kˆ × Uk + g H ∇(h k − h k ) + r Uk /H = ρk ,
(5.2.7)
iωk h k + ∇ · Uk = σk ,
(5.2.8)
where ρk and σk are dynamical misﬁts or residuals. Note that the bathymetry H only appears in the momentum equation, so the continuity equation should be very accurate. The boundary conditions are Uk · nˆ ∼ =0
at coasts,
hk ∼ = h Bk
at open boundaries.
(5.2.9)
We refrain from introducing symbols for the boundary residuals; instead we shall express (5.2.7)–(5.2.9) compactly as SU = F + τ ,
(5.2.10)
where S comprises the linear dynamical operators and linear boundary operators, F includes the tgf and boundary forcing, while τ includes the dynamical and boundary residuals.
5.2 Ocean tides
131
The crossover data involve the linear measurement functional Lm [U] = h Xm , Tm(2) − h Xm , Tm(1) ,
(5.2.11)
where Xm is a crossover location, while Tm(1) and Tm(2) are the times of distinct passes. Note that Lm selects only the elevation h, and evaluates it at certain times. In terms of complex harmonic amplitudes, the functional becomes Lm [U] =
e
K
(2) (1) h k (Xm ) eiωk Tm − eiωk Tm ,
(5.2.12)
k=1
for 1 ≤ m ≤ M. In compact form, the data are d = L[U] + .
(5.2.13)
Note that the ﬁeld U has values in C3K , while d belongs to R M .
Exercise 5.2.2 Devise the measurement functionals for data types (i)–(iv) in §5.2.3. Explain why representers for altimetric crossover data can be constructed using representers for tide gauge data. Finally, the penalty functional for inversion of the LTE and T/P data is (Egbert et al., 1994; Egbert and Bennett, 1996) J [U] = τ ∗ ◦ Wτ ◦ τ + ∗ w,
(5.2.14)
where ∗ denotes the transposed complex conjugate vector. Note that we should choose Wτ = C−1 τ , where Cτ is the covariance of residuals at different places and at different frequencies, and we should choose w = C−1 where C is the covariance of measurement errors.
5.2.6
Choosing weights: scale analysis of dynamical errors
Before proceeding with the mathematical task of minimizing the penalty functional J , let us take a ﬁrst look at the choice of weights in J . As these will be the inverses of error covariances, consider ﬁrst some scale estimates of dynamical errors. (i) The dynamics are linearized. Also, we have analyzed the ﬁelds harmonically, thus ∂h = iωh at frequency ω. Let us assume that ∂∂hx ∼ κδh, where κ is some ∂t wavenumber and δh is the rough magnitude of h. The balance between local accelerations and pressure gradients (5.2.7) may be expressed as ωδu ∼ gκδh.
(5.2.15)
5. The ocean and the atmosphere
132
The balance between the local rate of change of sea level and the convergence of volume ﬂux in (5.2.8) is ωδh ∼ κ H δu.
(5.2.16)
Hence ω2 ∼ g H κ 2 ,
κ∼
ω , c
(5.2.17)
1
where c = (g H ) 2 , and δu ∼
+ g , 12 ωδh δh δh = c . ∼ κH H H
(5.2.18)
Now compare the local acceleration and momentum advection in (5.2.7): ωδu : κ(δu)2 .
(5.2.19)
These are in the ratio 1:
κδu , ω
or 1 :
δu , c
or 1 :
δh . H
(5.2.20)
The linearization error in the continuity equation (5.2.8) is also (δh/H ). In deep water H ∼ 5000 m and δh ∼ 0.2 m, so linearization is highly accurate. Note 1 that c = (g H ) 2 ∼ 200 m s−1 , so δu ∼ 0.008 m s−1 . (ii) The pressure gradients in (5.2.7) are derived from the hydrostatic balance (not shown). Using the threedimensional incompressibility condition (not shown), we may deduce that the scale of the vertical velocity is δw ∼ κ H δu, hence the comparison of local vertical accelerations to the gravitational acceleration is ωδw : g,
or ωκ H δu : g,
(5.2.21)
or κ 2 H 2 (δu/c) : 1.
(5.2.22)
or cκ 2 H δu : g,
So the hydrostatic balance is extremely accurate for smallamplitude (δh H ), long waves (κ H 1) in deep water. The dynamics are “shallow” in the sense that κ H 1. (iii) A crude estimate of numerical accuracy is made by comparing the horizontal grid spacing x to the length scale κ −1 = cω−1 . For solar semidiurnal tides, −1 1 ω = 2π = (2π/43 200) s−1 ∼ d = 1.4 × 10−4 s−1 , 2 so κ −1 ∼ 200 m s−1 /(1.4 × 10−4 s−1 ) ∼ = 1.4 × 106 m = 1400 km. Thus, if x = 0.5◦ ∼ = 50 km, and the numerics are secondorder accurate, then
5.2 Ocean tides
133
truncation errors are entirely negligible. Tidal diffraction at peninsulae reduces the length scale signiﬁcantly. It is also common practice to reduce grid spacing 1 in shallow water according to the rule x ∝ H 2 .
Exercise 5.2.3 Justify the shallowwater grid rule given above.
(iv) The mean depth H (x) is commonly taken from the US Navy’s ETOP95 bathymetry, which is available at NCAR. These data are very doubtful at high latitudes. In the deep North Paciﬁc we can only guess that the error is about 100 m in 5000 m, or 2%. There are known to be far greater errors in, for example, the Weddell Sea. (v) We have adopted the crude drag law: iωu · · · = · · · − r u/H, where r = 0.03 m s−1 . It is common practice to replace r/H with r/ max[H, 200 m], in order to avoid excessive drag over the continental shelves. These drag formulae are usually tuned so that the tidal solutions are in reasonable agreement with data. Nevertheless, such drag laws are crude parameterizations, so it is prudent to assume that they are 100% in error. However, the drag is a very small part of the momentum balance in deep water. (vi) The rigid boundary condition is simply H u · nˆ = 0,
(5.2.23)
where nˆ is normal to the boundary. The question arises: where is the boundary? In a numerical model the precision of location is no smaller than x, so the error in (5.2.23) is of the order of x
∂ ˆ ∼ xκ H δu · n. ˆ (H u · n) ∂n
(5.2.24)
The relative error in (5.2.23) is therefore ∼xκ. If we assume that 1 κ ∼ ω(g H )− 2 , H ∼ 100 m, g ∼ 10 m s−2 and ω = S2 1.4 × 10−4 s−1 , then κ∼ = 0.5 × 10−5 m−1 . So if x = 0.5◦ ∼ = 50 km, then xκ ∼ = 0.25.
(5.2.25)
The relative error in (5.2.23) is 25%! The depth would have to increase to 10 km in order for xκ to be as small as 2.5% (given x = 0.5◦ ). So rigid boundary conditions are signiﬁcant sources of error in numerical tidal models. The solution in midocean may not be sensitive to this error source, as the basin resonances are very broad. That is, the coastal irregularity itself ensures a ﬁne spectrum of seiche modes. Finally, the highresolution Finite Element Model (FEM) for global tides developed at the Institute for Mechanics in Grenoble, France is the best forward model yet developed (Le Provost et al., 1994).
5. The ocean and the atmosphere
134
In summary, linearization and truncation errors in the continuity equation are negligible. Bathymetric errors and drag errors in the momentumtransport equations should be admitted, while rigid boundary conditions are signiﬁcantly in error.
The formalities of minimization
5.2.7
Let us set aside our preliminary discussion of model errors, and make some notes on the formalities of minimization. The penalty functional (5.2.14) is ∗ −1 J [U] = τ ∗ ◦ C−1 τ ◦ τ + C
(5.2.26)
∗ −1 ≡ (SU − F)∗ ◦ C−1 τ ◦ (SU − F) + (d − L[U]) C (d − L[U]).
(5.2.27)
Setting the ﬁrst variation of J to zero yields 1 ˆ ∗ −1 ˆ ˆ δ J = (SδU)∗ ◦ C−1 τ ◦ (SU − F) − L[δU] C (d − L[U]). 2 The vanishing of the coefﬁcient of δU∗ yields 0=
(5.2.28)
ˆ S† Λ = L[δ]∗ C−1 (d − L[U]),
(5.2.29)
SUˆ = F + Cτ ◦ Λ.
(5.2.30)
where
Note that S and the adjoint operator S† include the dynamics and the boundary conditions.
Exercise 5.2.4
Derive (5.2.29), (5.2.30) in detail.
Let us now examine L for TOPEX/POSEIDON crossover data (T/P XO data): Lm [U] = h Xi , T j(2) − h Xi , T j(1) , (5.2.31) where 1 ≤ m = m(i, j) ≤ M. The Xi for 1 ≤ i ≤ I are the XO locations; the T j(1,2) for 1 ≤ j ≤ J are the XO times. In terms of tidal constituents we have, from (5.2.12): Lm [U] =
e
K
(2) (1) h k (Xi ) eiωk T j − eiωk T j .
(5.2.32)
k=1
So it sufﬁces to calculate representers for h k (Xi ) for 1 ≤ i ≤ I and 1 ≤ k ≤ K . Then we can synthesize the representers for the (Xi , T j(1,2) ) XO difference. This is very useful. There are only 1 × 104 XO points but by 9/99 there had been approximately 258 tenday repeattrack orbit cycles, or about 1.8 × 106 XO data. According to the above harmonic analysis, we need only compute K × 104 representers (K is usually 4 or 8). How else might we reduce the computations? Inspection of reasonably accurate solutions of forward tidal models indicates that the XO coverage is unnecessarily dense,
5.2 Ocean tides
135
for observing tides. In the open ocean, adequate coverage is obtained with every third XO in each direction. Thus we may reduce the number of representers by nearly a factor of ten. Finally, a cheap preliminary calculation of all the remaining representers, using a coarse numerical grid, permits an array mode analysis (see §2.5). The analysis shows that a further reduction by a factor of about four is appropriate. In conclusion, about 4000 real representers are needed. They may be ﬁtted to the 1.8 × 106 data values, however.
5.2.8
Constituent dependencies
It might be inferred from the preceding discussion that the representers at different tidal frequencies may be calculated independently. In general, this is not the case. The representer adjoint variables obey †
Sk αk = δ(x − ξ)ˆe3
(5.2.33)
for 1 ≤ k ≤ K , where eˆ 3 = (0, 0, 1, 0)T , and so may be calculated separately. However the representers obey Sk rk =
K
Cτkl ◦ αl
(5.2.34)
l=1
for 1 ≤ k ≤ K , and in general the LTE error covariance is not diagonal with respect to k and l. Nevertheless, we may reasonably assume that errors for semidiurnal constituents are independent of those for diurnal constituents. The tidal inverse problem involves immense detail, because so much is known about the structure of the tidegenerating force.
5.2.9
Global tidal estimates
Estimating global ocean tides using hydrodynamic models and satellite altimetry is formulated as an inverse problem in Egbert et al. (1994). The altimetric data are being inverted in order to ﬁnd errors in the drag law and bathymetry, especially in the deep ocean. Linear dynamics and linear measurement functionals sufﬁce. The time dependence involves few degrees of freedom and those are highly regular, pure harmonic in fact. The number of crossover data and hence the number of representers is very large (and still growing, after eight years), yet their number can be reduced by obvious and reasonable subsampling strategies (for example, every third crossover in each direction), and by a priori array assessment based on economical computation of representers on a coarser numerical grid. The eventual set of decimated and rotated representers may still be ﬁtted closely to the entire data set, however. Best of all, the challenge of a real, large and important problem led (Egbert, personal communication) to the indirect representer method, outlined in §3.1.3 and applied to real
136
5. The ocean and the atmosphere
data in Egbert et al. (1994; hereafter referenced as EBF). This tidal solution and others have been extensively reviewed (Andersen et al., 1995; Le Provost, 2001; Le Provost et al., 1995; Shum et al., 1997). The solutions were tested with independent tide gauge data. All agreed to within a few centimeters, but the EBF inverse solution (“TPX0.2”) did not perform as well as empirical ﬁts to the altimetry (Schrama and Ray, 1994), nor as well as a ﬁniteelement forward solution of the Laplace Tidal Equations obtained by a team in Grenoble (Le Provost et al., 1994). The inverse solution was in effect an empirical ﬁt to the altimetry using a few thousand representers, whereas the other empirical ﬁts used around one hundred thousand degrees of freedom. Schrama and Ray (1994) chose the highresolution Grenoble ﬁniteelement solution as the prior, or ﬁrstguess for their empirical ﬁt. The prior for the EBF inverse was a ﬁnitedifference solution of the Laplace Tidal Equations on a relatively coarse grid. A striking and conﬁdenceenhancing aspect of the inverse solution was its relative smoothness, which it owed to its parsimony or few degrees of freedom. The Grenoble ﬁniteelement solution had very ﬁne resolution in shallow seas, where it excelled. The EBF inverse solution was based on representers for crossovers in deep water only. Driven by the tidegenerating force and tidal data at few basin boundaries, the ﬁniteelement model is almost a pure mechanical theory and so its success is all the more impressive. More recent implementations (Le Provost, personal communication) have no basin boundaries, that is, the domain is the global ocean and so no tidal data are needed to close the solution. Nevertheless, the tidal solutions are quite accurate. This is a remarkable technical and scientiﬁc achievement, surely the most successful theory in geophysics and one of the most successful in all of physics. The ﬁniteelement model is limited principally by inaccurate bathymetry and by incomplete parameterizations of drag. It has recently been reformulated as an inverse model, and solved with representers computed by ﬁniteelement methods (Lyard, 1999). The latest tidal solutions of various type, now based on eight or more years of altimetry and reﬁned orbit theories, are believed to agree to well within observational errors (e.g., Egbert, 1997). A new independent trial is underway at the time of writing (October 2001). The most recent ﬁnitedifference inverse solution (TPX0.4) uses approximately 4 × 104 real valued representers, including many in shallower seas (Egbert and Ray, 2000). A global plot of coamplitude and cophase lines may be found at www.oce.orst.edu/po/research/tide/global.html. A unique feature of the inverse tidal solutions is the availability of maps of residuals in the equations of motion – the Laplace Tidal Equations. A global plot of the average, per tidal cycle, of the rate of working by the dynamical residuals for the principal lunar semidiurnal constituent M2 of TPX0.4, is shown in Fig. 5.2.7. Negative values indicate that the tides are losing energy. The largest losses do not occur in regions of the strong boundary currents of the general circulation, such as the Gulf Stream, but instead along the ridges and other steep topography. These errors may be due to the somewhat simpliﬁed parameterizations of earth tide and load tide, to unresolved topographic waves or to internal tides. The net loss is a delicate balance involving work done by residuals, by a model bottom drag and by the moon.
5.2 Ocean tides
137
60 40 20 0 ⫺20 ⫺40 ⫺60 0
50 ⫺25
100 ⫺20
150 ⫺15
200
250
⫺10
300
⫺5
350 0
W/m
2
Figure 5.2.7 Percycle average rate of working of the M2 dynamical residuals in TPX0.4 on the M2 tide, in units of W m−2 . Negative values indicate that the M2 tide is losing energy (after Egbert, 1997).
60˚ 200 kW/m
30˚
0˚
30˚
60˚ 60˚
120˚
180˚
120˚
60˚
0˚
Figure 5.2.8 Flux of total mechanical energy for the linear semidiurnal tidal constituent M2 , based on the inverse model TPX0.4. Note especially the convergence into regions of significant tidal dissipation: for example, the North West Australian shelf, Micronesia/Melanesia and the European shelf (Egbert and Ray, 2001: Estimates of M2 tidal energy dissipation from TOPEX/POSEIDON altimeter data, c 2001 American Geophysical Union, reproduced by J. Geophys. Res., in press. permission of American Geophysical Union).
The various highly accurate tidal solutions are leading to refined estimates of the dissipation of the energy input to the ocean by the tidegenerating force (Lyard and Le Provost, 1997; Le Provost and Lyard, 1997; Egbert and Ray, 2000, 2001). These estimates show that tidal dissipation can provide about 50% of the 2TW of power believed to sustain the meridional overturning circulation, the other 50% being provided by the wind (Wunsch, 1998). A map of energy flux vectors for the tides is shown in Fig. 5.2.8. Some of this power is being produced by the dynamical residuals. Note
138
5. The ocean and the atmosphere
that there are strong convergences and divergences in deep water, as well as ﬂuxes towards the marginal seas. These deepwater convergences and divergences almost exactly balance the work done by the moon.
5.3
Tropical cyclones (1). Quasigeostrophy; track predictions
5.3.1
Generalized inversion of a quasigeostrophic model
The linear “toy” ocean model of §1.1 and the linear Laplace Tidal Equations of §5.2 are not representative of ocean circulation, which has nonlinear dynamics and thermodynamics. The generalized inverse of a nonlinear quasigeostrophic model is now deﬁned and the Euler–Lagrange equations are derived. The latter equations are also nonlinear, so the linear representer algorithm can only be applied iteratively. Two iteration schemes are introduced; a more extensive analysis is provided in §3.3.3. The formulation of an hypothesis for the dynamical errors, which difﬁcult subject was broached §5.2, is considered further here. Finally, implementation of the representer algorithm is discussed in some detail.
5.3.2
Weak vorticity equation, a penalty functional, the Euler–Lagrange equations
Let us formulate weak quasigeostrophic dynamics entirely in terms of the streamfunction ψ: ∂ 2 ∂(ψ, ∇ 2 ψ + f ) ∇ ψ+ = τ, ∂t ∂(x, y)
(5.3.1)
where τ is the residual in the quasigeostrophic vorticity equation. We shall also specify a weak initial condition for ψ: ψ(x, 0) = ψ I (x) + i(x),
(5.3.2)
where i is the initial residual. A weak condition for ψ on the boundary B of the simplyconnected domain D is ψ(x, t) = ψ B (x, t) + b(x, t),
(5.3.3)
where x lies on B, and b is the boundary streamfunction residual. We shall weakly specify the relative vorticity all around B: ∇ 2 ψ(x, t) = ξ B (x, t) + z(x, t), where x lies on B and z is the boundary vorticity residual. See Fig. 5.3.1.
(5.3.4)
5.3 Tropical cyclones (1)
139
Figure 5.3.1 Planar domain D with open boundary B. Streamlines cross B at least twice, or not at all. What can be said about particle paths?
Now (5.3.1) is equivalent to Dζ ∂ζ = + u · ∇ζ = τ, Dt ∂t
(5.3.5)
where ζ ≡ ∇ 2 ψ + f . It follows that if τ is prescribed, then ζ is determined by integrating (5.3.5) along a particle path from an initial position either inside D or on the boundary B. If any particle path exits D in the time interval of interest, then (5.3.4) overdetermines ζ . However we shall “adjust” the residuals τ , i, b and z, so that a continuous solution is obtained for ζ , and hence for ψ. More precisely, we shall seek ψ yielding a smooth weightedleastsquares bestﬁt to (5.3.1)–(5.3.4). A suitable penalty functional is (Bennett and Thorburn, 1992): J [ψ] = τ • Cτ−1 • τ + i ◦ Ci−1 ◦ i + b ∗ Cb−1 ∗ b + z ∗ C z−1 ∗ z + Jd ,
(5.3.6)
where Jd is a penalty for misﬁts to data within D. Note that T •≡ D
0
T
dt da, ◦ =
da, ∗ = D
dt ds,
(5.3.7)
0 B
and all integrations are on a surface of constant pressure p. If ψˆ is a local extremum of J , then 1 ˆ = δτ • Cτ−1 • τˆ + δψ ◦ Ci−1 ◦ iˆ δJ [ψ] 2 1 + δψ ∗ Cb−1 ∗ bˆ + δξ ∗ C z−1 ∗ zˆ + δJd = 0. 2
(5.3.8)
We shall manipulate the ﬁrst two terms in detail, leaving the boundaries as an exercise. ˆ ≡ Cτ−1 • τˆ . Then Deﬁne λ T ˆ = δτ • λ 0
D
∂ ∂(ψ, ∇ 2 ψ + f ) ˆ dt da δ ∇ 2 ψ + δ λ(x, t). ∂t ∂(x, y)
(5.3.9)
140
5. The ocean and the atmosphere
The ﬁrst term in (5.3.9) is easily manipulated: ∂ 2 ∂ 2 ˆ ˆ δ ∇ ψ λ= ∇ δψ λ ∂t ∂t ∂ ˆ 2 ∂ ˆ λ ∇ 2 δψ. = ( λ∇ δψ) − ∂t ∂t
(5.3.10)
Integrating the ﬁrst term in (5.3.10) over time yields T ˆ 2 δψ]t=0 [ λ∇ .
(5.3.11)
ˆ 2 δψ = ∇ · ( λ∇δψ ˆ ˆ + δψ∇ 2 λ. ˆ λ∇ − δψ∇ λ)
(5.3.12)
Now
If the area integral of terms proportional to δψ(x, T ) vanishes, for arbitrary values of the latter, then ˆ = 0 at t = T. ∇2 λ
(5.3.13)
ˆ + Ci−1 ◦ (ψˆ − ψ I ) = 0 at t = 0. −∇ 2 λ
(5.3.14)
Similarly, we infer that
The second term in (5.3.10) is ˆ ˆ ˆ ∂ ˆ ∂λ ∂λ ∂λ 2 − λ ∇ δψ = −∇ · ∇δψ − δψ∇ − δψ∇ 2 . ∂t ∂t ∂t ∂t
(5.3.15)
The second term in (5.3.15) will be used shortly. Consider now the variation of the Jacobian in (5.3.9): 2 2 ˆ 2 ˆ 2 ˆ ∂(ψ, ∇ ψ + f ) = λ ˆ ∂(δψ, ∇ ψ + f ) + λ ˆ ∂(ψ, δ∇ ψ) + O( λ(δψ) ˆ λδ ) ∂(x, y) ∂(x, y) ∂(x, y)
. . . = − δψ
(5.3.16)
ˆ ∇ 2 ψˆ + f ) ˆ ψ) ˆ ∂( λ, ∂( λ, + ∇2 ∂(x, y) ∂(x, y)
+ divergence terms.
(5.3.17)
So, by requiring the coefﬁcient of δψ(x, t) to vanish, we recover from (5.3.8), (5.3.15) and (5.3.17) the Euler–Lagrange equation −
ˆ λ) ˆ ˆ ∇ 2 ψˆ + f ) 1 δJd ∂ 2ˆ ∂(ψ, ∂( λ, ∇ λ − ∇2 = − , ∂t ∂(x, y) ∂(x, y) 2 δψ
(5.3.18)
where the last term is a linear combination of measurement kernels. Equation (5.3.18) may be formally rewritten as ˆ ∂λ δJd 1 ˆ = ∇ −2 µ − − ∇ · (uˆ λ) , (5.3.19) ˆ · ∇ζˆ − ∇ −2 ∂t 2 δψ
5.3 Tropical cyclones (1)
141
ˆ ˆ ˆ ˆ where uˆ = − ∂∂ψy , ∂∂ψx and µ ˆ ≡ − ∂∂ yλ , ∂∂ xλ . Note that ∇ · uˆ = ∇ · µ ˆ = 0. The form (5.3.19) looks like the “adjoint” of the total vorticity conservation law ∂ζ + u · ∇ζ = τ, ∂t
(5.3.20)
but for the emergence of a new term on the rhs of (5.3.19) arising from the variation of the advecting velocity u: δτ =
∂ δζ + u · ∇δζ + (δu) · ∇ζ. ∂t
(5.3.21)
No such term arose in our “toy” model ∂u + c ∂∂ux = τ , since the phase velocity c was ∂t ﬁxed. In particular c did not depend upon the state u.
Exercise 5.3.1 Derive the boundary conditions that accompany the variational equation (5.3.18) or (5.3.19). Which boundary condition goes with which equation?
Iteration schemes; linear Euler–Lagrange equations
5.3.3
The Euler–Lagrange system (5.3.1) and (5.3.18), with attendant initial and boundary conditions, is nonlinear. The system is coupled through the data term (δJd /δψ) in (5.3.18), through advection on the lhs of (5.3.18) and through the other term on the rhs of (5.3.18). It is also coupled through the boundary conditions. A simple iteration scheme breaks the coupling completely: calculate a sequence {ψˆ n , λn }∞ n=1 such that ∂ 2 ˆ n ∂(ψˆ n , ∇ 2 ψˆ n + f ) ∇ ψ + = C τ • λn , ∂t ∂(x, y) −
∂ 2 n ∂(ψˆ n−1 , λn ) ∂(λn−1 , ∇ 2 ψˆ n−1 + f ) 1 δJdn−1 ∇ λ − ∇2 = − . ∂t ∂(x, y) ∂(x, y) 2 δψ
(5.3.22) (5.3.23)
These equations may be solved by integrating (5.3.23) backwards, and (5.3.22) forwards. Note that (5.3.22) is nonlinear in ψˆ n , but the broken coupling eliminates the need for representers! However, such a sequence always seems to diverge. An alternative iteration scheme is: ∂ 2 ˆ n ∂(ψˆ n−1 , ∇ 2 ψˆ n + f ) ∇ ψ + = C τ • λn , ∂t ∂(x, y) −
∂ 2 n ∂(ψˆ n−1 , λn ) ∂(λn−1 , ∇ 2 ψˆ n−1 + f ) 1 δJdn ∇ λ − ∇2 = − . ∂t ∂(x, y) ∂(x, y) 2 δψ
(5.3.24) (5.3.25)
This system is linear, but coupled: note that the data term in (5.3.25) is evaluated with ψˆ n . It is the Euler–Lagrange system for a linear dynamical model, advected by uˆ n−1 . There is a ﬁrstguess forcing Cτ • λnF , where λnF is the response of the lhs of (5.3.25) to the ﬁrst term on the rhs. The system may be solved using representers. The sequence
142
5. The ocean and the atmosphere
converges in practice if the Rossby number is moderate. There are some theorems about convergence in doublyperiodic domains: the sequence is bounded and so must have points of accumulation or cluster points, but not necessarily unique limits. There is numerical evidence of the sequence cycling, presumably between cluster points. A third iteration scheme is described in §3.3.3.
5.3.4
What we can learn from formulating a quasigeostrophic inverse problem
A quasigeostrophic inverse model offers some especially clear problems in error estimation. Also, there are special opportunities for reliable estimation of these errors. To the extent that the situation is not representative of Primitive Equation inverse models, one might regard quasigeostrophic inversion as a curiosity but, like the uniquely elegant tidal inverse problem, the quasigeostrophic inverse problem offers valuable experience.
5.3.5
Geopotential and velocity as streamfunction data: errors of interpretation
There are errors of interpretation in certain streamfunction data. Consider geopotentials φ and horizontal velocities u measured by radartracking of high altitude balloons, or by sonartracking of deeply submerged ﬂoats. The quasigeostrophic state variable is the streamfunction ﬁeld ψ. We must relate φ and u to ψ. The geostrophic approximation is f kˆ × u ∼ = −∇φ,
(5.3.26)
where the Coriolis parameter is a function of latitude: f = f (y). We have assumed that ∇ · u ∼ = 0 and that there is a streamfunction for ψ, so (5.3.26) becomes − f ∇ψ ∼ = −∇φ.
(5.3.27)
Ignoring variations in f leads to the “poor man’s balance equation” f 0 ψ = φ,
(5.3.28)
where f 0 = f (y0 ) for some latitude y0 . Hence geopotential data may be used as approximations to streamfunction data. Also, velocity data may be used as approximations to streamfunctiongradient data: ∇ψ = −kˆ × u.
(5.3.29)
Let us begin to estimate the errors in (5.3.28) and (5.3.29). If L is a horizontal length scale and U is a velocity scale, then the local acceleration neglected in (5.3.26) has the scale U 2 L −1 . The Coriolis acceleration retained in (5.3.26) has the scale f 0 U , so the relative errors in (5.3.26) scale as the Rossby number Ro ≡ fU0 L . We shall assume for
5.3 Tropical cyclones (1)
143
simplicity that variations in f are smaller than Ro f 0 . Then (5.3.26) is kˆ × u = − f 0−1 ∇φ + O(U Ro),
(5.3.30)
hence ∇·u= O
Ro
U L
.
(5.3.31)
In general, for any u there is a streamfunction ψ and a velocity potential χ such that u = kˆ × ∇ψ + ∇χ ≡ uψ + uχ ,
(5.3.32)
thus ξ ≡ kˆ · ∇ × u = ∇ 2 ψ,
δ ≡ ∇ · u = ∇ 2χ .
(5.3.33)
We conclude from (5.3.31) that χ is O(Ro U L) and hence (5.3.29) is accurate to O(Ro U ), while (5.3.28) is accurate to O(Ro f 0 U L). In summary, the “theoretical” relative errors in the data are O(Ro), where Ro ≡ U/( f 0 L). For Gulf Stream meanders in the ocean, U = 1 m s−1 (= 2 knots), L = 105 m and f 0 = 10−4 s−1 , so Ro = 0.1. For middlelevel synopticscale weather systems in the atmosphere, U = 30 m s−1 and L = 106 m, so Ro = 0.3. In the preceding analysis, the estimates of neglected local accelerations were based on the values L and U representative of the synopticscale circulation of interest. For consistency, all ﬁelds should be lowpass ﬁltered prior to sampling, in order to suppress smallerscale motions such as internal waves. If, as is often unavoidable, the smoothing is inadequate, then the data will be contaminated with aliased signals. This contamination can be substantial, exceeding for example the estimate O(RoU ) for errors in (5.3.29). (I am grateful to Dr Ichiro Fukumori for a discussion of this point. AFB)
5.3.6
Errors in quasigeostrophic dynamics: divergent ﬂow
Estimating the dynamical errors in a quasigeostrophic model is particularly instructive, as we have closed analytical forms for many sources of error. Recall again the momentum balance for the Primitive Equations: ∂u ∂ + (u · ∇)u + ω u + f kˆ × u = −∇φ. ∂t ∂p
(5.3.34)
Taking the curl at constant pressure yields ∂ξ ∂ξ ∂u + (u · ∇)ξ + ω + kˆ · ∇ω × + ( f + ξ )δ + βv = 0, ∂t ∂p ∂p
(5.3.35)
144
5. The ocean and the atmosphere
where ξ = kˆ · ∇ × u, δ = ∇ · u and β ≡ d f /dy. Splitting u into a solenoidal part uψ and an irrotational part uχ (see (5.3.31)) leads to a split for (5.3.35): ∂ξ ∂ξ ˆ ∂u + (uψ · ∇)ξ + βvψ = −(uχ · ∇)ξ − ( f + ξ )δ − βvχ − ω −k·∇ ω× ≡ τ. ∂t ∂p ∂p (5.3.36) That is, ∂ 2 ∂(ψ, ∇ 2 ψ + f ) ∇ ψ+ = τ, ∂t ∂(x, y)
(5.3.37)
where we have an explicit form for τ in terms of resolvable ﬁelds. That is, given archives of gridded ﬁelds of u(x, p, t), we may evaluate τ on the grid, and hence estimate its mean Eτ and covariance Cτ . The most difﬁcult part is calculating ω reliably. We could use the conservation of mass: ∂ω = −∇ · u, ∂p subject to ω → 0 as p → 0, or we could use the conservation of entropy: ˙ −1 Q ∂η ∂η ω= − − u · ∇η , T ∂t ∂p
(5.3.38)
(5.3.39)
where Q˙ is the heat source per unit mass and T is the absolute temperature. Note that in order to calculate η and T via the equation of state, we need the other thermodynamic state variables such as ( p, ρ, q) in the atmosphere, or ( p, ρ, S) in the ocean. We may dispense with ρ if T has been measured or is otherwise available on the grid. There are opportunities to make similar direct estimates of dynamical errors in other “reduced” models, such as balanced models, and the Cane–Zebiak coupled model (Zebiak and Cane, 1987). However, there are additional dynamical errors in all these reduced models, owing to unresolved stresses. The additional errors may exceed the resolvable errors.
5.3.7
Errors in quasigeostrophic dynamics: subgridscale ﬂow, second randomization
We shall consider the unresolved stresses, in the context of the quasigeostrophic vorticity equation ∂ξ + u · ∇ξ + βv = 0, ∂t
(5.3.40)
where ξ = ∇ 2 ψ and u = kˆ × ∇ψ. Note that the subscript “ψ” on u is now dropped. In practice we can only calculate with (5.3.40) on a grid having some ﬁnite resolution in space and time. Yet we know from observations and from instability theory that (5.3.40) possesses solutions that have inﬁnitesimally ﬁne structure of signiﬁcant amplitude.
5.3 Tropical cyclones (1)
145
We try to separate the coarse and ﬁne structures using the abstraction of an ensemble of ﬂows having a mean (with only the coarse scales), and variability (with only the ﬁne scales). That is, ξ = ξ + ξ , where ξ = ξ and ξ = 0. In practice we can only approxi( ), but mate the ensemble average (denoted by ( ) here) using a space or time average then ( ξ˜ ) = ξ˜ . We shall ignore this very important issue here (see Ferziger, 1996 for an excellent discussion) and assume that ( ) may be estimated with adequate accuracy. Only the mean ﬁeld being of interest, it would be desirable to replace the “detailed” vorticity equation (5.3.40) with an equation for ξ , u and ψ. Averaging (5.3.40) yields ∂ ξ + u · ∇ξ + βv = 0. ∂t
(5.3.41)
Now for any a and b, ab = (a + a )(b + b ) = (ab + ab + a b + a b ) = a b + a b + a b + a b
(!)
= ab + a0 + 0b + a b = ab + a b .
(5.3.42)
Thus u · ∇ξ = u · ∇ξ + u · ∇ξ
(5.3.43)
∂ξ + u · ∇ξ + βv = τ ≡ −∇ · (u ξ ), ∂t
(5.3.44)
and (5.3.41) becomes
where we have used ∇ · u = 0. So there is another candidate for the residual τ in the mean dynamics: the divergence of the mean “eddyﬂux” of relative vorticity. Finding a formula for such ﬂuxes in terms of ﬁrst moments (that is, in terms of ξ , u or ψ) is the turbulence problem. It remains unresolved. However, we may use (5.3.44) to constrain the circulation, provided we can put bounds on τ . At this point the fast talk begins. Realizing that even the smoothed ﬁelds ﬂuctuate considerably, we may regard τ as a random ﬁeld with a prior mean and variance (prior to assimilating data). Generally we neglect the new mean Eτ for τ (or else model it with a diffusion law, for example), and struggle to make scale estimates for the variability in τ . For example, if for the eddies u ∼ U and x ∼ l, we might be tempted to assume τ ∼ U 2l −2 . This is usually excessive; the length scale L of the (smoothed) eddyﬂux u ξ is much greater than the length scale l of the eddies themselves. That is, τ ∼ cU 2 L −1l −1 , where c 1 is the magnitude of the correlation coefﬁcient between u and ξ . The decorrelation length scale D for τ presumably lies in the interval l < D < L, while the decorrelation time T lies in the range (l/U ) < T < (L/U ). In the jet stream or ocean boundary currents, on
146
5. The ocean and the atmosphere
the other hand, l ∼ L. For the weakly homogeneous case (l L), however, we might hypothesize that c2 U 4 x − y2 t − s2 E(τ (x, t)τ (y, s)) = Cτ (x, t, y, s) = 2 2 exp − . (5.3.45) − L l D2 T2 One might reasonably feel uncomfortable at this point, attempting to constrain a circulation estimate with such a speculative hypothesis. Indeed, the “second randomization” of τ is a na¨ıve abstraction of the hopedfor scale separations in the ﬂuctuations in τ . One should recall that the conventional forward model is merely a circulation estimate based on the hypothesis that τ ≡ 0. This is the one hypothesis that we know immediately to be wrong. We could abandon the concept of an ensemble of mean vorticity ﬂuxes, and just manipulate τ as a control that guides the state towards the data. The Euler–Lagrange equations of the calculus of variations enable the manipulations, once a penalty functional has been prescribed. The difﬁculty lies in the choice of weights. Probabilistic choices (inverses of covariances) are conceptually shaky. Yet the prospect of an ocean model as a testable hypothesis is so appealing. It was established in §2.2 that generalized inversion is equivalent to optimal interpolation in space and time. The former requires the dynamical error covariance Cτ ; the latter requires the circulation or state covariance such as Cψ . Which is the easier to specify a priori? We anticipate that ψ is nonstationary, anisotropic and signiﬁcantly inhomogeneous. The components of multivariate circulation ﬁelds will be jointly covarying. On the other hand, it is plausible that the dynamical residuals in unreduced models are the result of smallscale processes that are locally stationary, isotropic and univariate. Then the generalized inverse constructs highly structured state covariances guided by the model dynamics, and by the morphology of the domain: the orography, or the bathymetry and coastline. 5.3.8
Implementation; ﬂow charts
The linear representer method is complicated. Its iterative application to a nonlinear quasigeostrophic model makes it even more complicated. Some general suggestions on implementation are in order. (i) Start with a simple, linear problem ﬁrst, such as the one described in §1.1–§1.3. The computing exercises at the end of this book provide numerical details. FORTRAN code is available from an anonymous ftp site: ftp.oce.orst.edu, cd/dist/bennett/class. (ii) A ﬂow chart for the “quasigeostrophic inverse” is given in Figs. 5.3.2 and 5.3.3. The latter ﬁgure shows in detail the hatched section in the former. These computations are manageable using a workstation. Your code should consist of a main program that calls many subroutines. These should include a single “backward integration” and a single “forward integration”. Preconditioned conjugate gradient solvers are widely available in subroutine libraries.
START n=1 ^u
o ≡ uI
initial vel(!)
^ un1 = u n1
adv. vel
Initial, boundary vort
Prior error cov’s
ζI , ζ B
ζF
ψB boundary str.
ψF
d data
Czz,Cφφ,Cθθ,Cττ,Cεε
n
n
Prior Est
hn = d (ψF ) n
^
ζn
ζB , ζ I
^ ψn , u n
ψB
? n= ∞
convergence
prior data error DATASPACE SEARCH ENGINE
prior Fn penalty
inverse ^
Cεε
YES
STOP
NO
Statistical Simulation ^ of posterior Cψψ
n = n +1
iterate
Figure 5.3.2 Generalized inversion of a regional quasigeostrophic model; indirect representer algorithm.
^u n1 (0)
bn = h n
priors
Adjoint & Homogeneous Forward Integration (k)
δn (k)
[δn(k)] measurement
Pbn =
(k+1)
bn
1
Pc preconditioner
∇jn = Pc Pnbn  Pc hn
preconditioned conjugate gradient solver
1
jn [bn] ≡ 2bnT (Pc Pnbn  Pc hn) (k)
k = k+1 hn
significance test ^ ? χ2 n
=
M
NO
^ (∞) bn
1
(k)
?
k=∞
1
convergence
YES
rep coeffts
^
Adjoint & Forced Forward Integration
Figure 5.3.3 Dataspace search engine.
ζn
5. The ocean and the atmosphere
148
(iii) The representer matrix should be tested for symmetry. The optimal values βˆ for the representer coefﬁcients, found by the gradient solver, should be compared ˆ to the values available a posteriori by measuring the inverse u(x, t): ˆ − d). βˆ = −w(L[ ˜ u]
(5.3.46)
(iv) Very large problems, such as those described in the following section, require very powerful computers having massive memory and disk. It is difﬁcult to offer further suggestions about implementations as each manufacturer provides a unique software development environment. The results presented in this chapter were obtained using Connection Machines. (v) For pseudocode, code on ftp sites and extensive details for implementation see Chua and Bennett (2001).
Track prediction
5.3.9
The intensity of tropical cyclones1 is controlled by the thermodynamics of the atmosphere and ocean together. Predicting the intensity requires a fully stratiﬁed coupled model. These are highly sensitive to the parameterization of heat exchange (Emanuel, 1999). Predicting the track of a typhoon, however, appears to be far simpler. The track is largely determined by “steering” winds, taken to be either an average over the ﬁelds in midtroposphere, or else the ﬁelds at 500 mb. In either case, the evolution of these winds can be represented for a short time (say, onethird of a synoptic time scale, or about a day) by singlelevel, quasigeostrophic dynamics. These simpliﬁed dynamics are nonetheless nonlinear, and so provide a relatively simple yet real and motivated ﬁrst test for timedependent variational assimilation in a nonlinear model. The formulation has been discussed in some depth already, so only data need be discussed here. Further details may be found in the ensuing references. The state variable for the quasigeostrophic model (5.1.38) is the streamfunction ψ = ψ(x, y, p, t); isobaric velocity u and vorticity ξ may be derived from it: see (5.1.33), (5.1.34). Observations through the entire depth of atmosphere were collected during a typhoon season by an international effort, the Tropical Cyclone ’90 or TCM90 experiment (Elsberry, 1990). These data were interpolated onto regular grids by the Australian Bureau of Meteorology Research Centre (Davidson and McAvaney, 1981). The BMRC tropical analysis scheme uses a threedimensional univariate statistical interpolation method. Vortex centers were inserted manually and synthetic proﬁles were used to generate “observations” for the statistical analyses (Holland, 1980). The gridded velocities were then partitioned into a rotational ﬁeld uχ satisfying kˆ · ∇ × uχ = 0, and a solenoidal ﬁeld uψ satisfying ∇ · uψ = 0. The gridded streamfunction ψ for the latter ﬁeld became the data for the quasigeostrophic assimilation. These streamfunction “data” were far from being direct measurements of the 1
Or typhoons, as they are known in the Paciﬁc (“hurricanes” in the Atlantic).
5.3 Tropical cyclones (1)
48
(t )
1100 2300 1100 1100
1100
(GMT)
(8/28)
(8/31)
(1990)
24
12
0
24
(8/29) (8/30)
149
Figure 5.3.4 Time line for generalized inversion of the tropical cyclone model. Initial data are available at t = −24 in the form of TCM90 analyses. Boundary data are available for −24 ≤ t ≤ 0, also from TCM90 analyses. For 0 ≤ t ≤ 48, the boundary data are provided by a global forecast of relatively coarse resolution. The inverse takes advantage of additional TCM90 data, at t = −12 hours and at t = 0 hours (as indicated by arrows). The dates refer to TC “Abe”.
atmosphere. They were further modiﬁed by a projection onto the leading ten empirical orthogonal functions (EOFs), which captured 94% of the variance. The time line for these derived data, and for the assimilationforecast episode is shown in Fig. 5.3.4. The gridded streamfunction data at t = −24 hours constitute the prior initial condition, ten EOF amplitudes were admitted at t = −12 and also at t = 0, and the smoothing or inversion interval was −24 ≤ t ≤ 0. The gridded inverse streamfunction at time zero, ˆ 0), became the initial condition for a forward integration or “forecast” that is, ψ(x, out to t = +48. The forecast was honest: boundary data for 0 ≤ t ≤ 48 were obtained from a global forecast model also starting at t = 0, rather than from archived analyses. Ten cases were considered; some involved the same typhoon in different stages of its life. Detailed results may be found in Bennett et al. (1993). From a scientiﬁc perspective the most interesting result is that the values of the 2 reduced penalty functional Jˆ were broadly in the range 20 ± 6, as expected for χ20 . Thus the hypothesized error covariances were consistent with the data. A diagnosis showed that the dynamical residuals and boundary vorticity residuals were negligible, so it was the hypothesized error covariances for the initial conditions and data that were consistent with the data. From a controltheoretic perspective the most interesting result is that there were sufﬁciently many degrees of freedom in the initial residuals at t = −24 to “aim” the model at the few data, without additional “guidance” from dynamical residuals for −24 ≤ t ≤ 0. From a forecasting perspective the most interesting result was the skill enhancement relative to other track prediction methods: see Fig. 5.3.5. Fortyeighthour track predictions based on variational assimilation over the preceding 24 hours were always superior to those based on either a carefully tuned “nudging” scheme and/or a purely statistical scheme (Bennett, Hagelberg and Leslie, 1992).
5. The ocean and the atmosphere
150
Improvement (%)
40
30
20
10
0 0
12
24
36
48
Hours
Figure 5.3.5 Tropical cyclone track predictions. Percentage improvement in (reduction of ) mean forecast error to 48 hours ahead, relative to climatology plus persistence (CLIPER). Diamonds: standard initialization; circles: initialization by generalized inversion. By convention the 7% difference at 48 hours, for example, is scaled with respect to the 70% residual yielding a “10% reduction”.
Tropical cyclones (2). Primitive Equations, intensity prediction, array assessment
5.4
The evolution of a tropical cyclone is a thermodynamic process. Quasigeostrophic dynamics assume that the stratiﬁcation remains close to the mean, and such is not the case in a tropical cyclone. The Primitive Equations (see §5.1) include laws for: (i) (ii) (iii) (iv)
conservation of mass in a fully compressible gas, (5.1.9); conservation of relative humidity q, (5.1.24); conservation of entropy η, (5.1.24); and an equation of state, (5.1.26), relating density ρ to η, to q and to the pressure p.
The state variable η may be replaced with the temperature T deﬁned by the com ∂η ∂η bined ∂η ﬁrst and second laws−1of thermodynamics: ∂ T p = C p /T , ∂ T v = Cv /T , = p/T , where v = ρ is the speciﬁc volume, while C p and Cv are respec∂v T tively the speciﬁc heats at constant pressure and volume. For adry, caloriﬁcally perfect γ +1 η/Cv (C p , Cv constant) ideal (η = Cv ln( pρ −C p/Cv )) gas,2 T = T0 ρρ0 e . Empirical corrections may be made for moisture: see, e.g., Wallace and Hobbs (1977). 2
Boltzmann’s equation leads directly to this deﬁnition of an ideal gas in terms of its entropy dependence, rather than in terms of the gas law p = RρT . The latter merely deﬁnes temperature. See, e.g., Chapman and Cowling (1970).
5.4 Tropical cyclones (2)
151
The full Primitive Equations may be found in Haltiner and Williams (1980, p. 17) or in the appendix to Bennett, Chua and Leslie (1996, hereafter BCL1; the associated Euler–Lagrange equations are also here in Appendix B). The vertical coordinate is not simply the pressure p as in §5.1, but Phillip’s sigmacoordinate: σ = p/ p∗ where p∗ is the pressure at the earth’s surface. The lower boundary for the atmosphere is conveniently located at σ = 1. A quadratic penalty functional for reconciling dynamics, initial conditions and data is also given in BCL1, along with (i) (ii) (iii) (iv)
the nonlinear Euler–Lagrange equations; the linearized Primitive Equations and Euler–Lagrange equations; the representer equations and the adjoint representer equations.
The linearized equations (ii)–(iv) enable an iterative solution of the nonlinear equations (i); each linear iterate may itself be solved by the indirect, iterative representer method described in §3.1. The “inner” or “data space” search was preconditioned in BCL1 using all representers calculated on a relatively coarse 128 × 64 × 9 global grid with twominute time steps, see Bennett, Chua and Leslie (1997, hereafter BCL2). The smallness of the time steps is due to the polar convergence of the meridians. The inverse was calculated on a relatively ﬁne 256 × 128 × 9 global grid with oneminute time steps. There were 4.4 × 108 grid points in a twentyfourhour smoothing interval, for about 2.6 × 109 gridded values of u, v, σ˙ , T , q, ln p∗ , etc. The coarsegrid preconditioner was only moderately effective owing to errors of interpolation from the coarse grid to the data sites. The latter were reprocessed cloud track wind observations (RCTWO) inferred from consecutive satellite images of middle and upperlevel clouds (Velden et al., 1992). Some of these observations are shown in Fig. 5.4.1. The observation period included tropical cyclone “Ed” near (113◦ E, 15◦ N) and Supertyphoon3 “Flo” near (130◦ E, 23◦ N). The RCTWO were available at t = −24, −18, −12 and 0 hours, and at 850 hPa, 300 hPa and 200 hPa for a total of M = 2436 vector components. The measurement errors for each component were assumed to be 3 m s−1 , 4 m s−1 and 4 m s−1 at the respective levels, uncorrelated from the other component of the same vector and from all other vectors elsewhere and at different times. The single inversion reported in BCL1 reduced the penalty functional from a prior value of 6432 to a posterior value of 4066. It may be concluded that the forward model and initial conditions (an ECMWF analysis) were very good, that the RCTWO only had moderate impact, and that the prior root mean square error should have been 30% larger. Given the difﬁculty in estimating the dynamical errors, such a conclusion is incontestable. Assimilation of the RCTWO did however have a useful impact on subsequent forecasts of meridional wind ﬁelds near “Flo”: see BCL1. Of greater interest here are the representers, for the Primitive Equation dynamics linearized 3
According to the Japanese Meteorological Agency.
170E
110E
150E
170E
170E
110E
130E
150E
170E
10S
10S
150E
0
0
0
130E
10N
10N
10N
110E
20N
20N
20N
10S
30N
30N
30N
40N
40N
300 hPa
10S
0
40N
200 hPa
130E
10S
10S
150E
0
0
130E
10N
10N
10N
110E
20N
20N
20N
110E
110E
130E
850 hPa
130E
850 hPa
150E
150E
170E
170E
Figure 5.4.1 Reprocessed cloud track wind observation vectors at 200 hPa, 300 hPa and 850 hPa, at t = −12 or 0000UTC on 16/IX/1990 (upper panels), and at t = 0 or 0000UTC on 16/XI/1990 (lower panels). RCTWO data are also available at t = −24, and at t = −18, for a total of 2436 scalar data (after Bennett, Chua and Leslie, 1996).
(t=0)
(t=12)
30N
30N
30N
40N
40N
300 hPa
40N
200 hPa
5.4 Tropical cyclones (2)
Figure 5.4.2 Zonal velocity ﬁelds at ( p = 200 hPa, t = 0) for representers of two RCTWO zonal components also at (200,0). Units are m2 s−2 , as for velocity autocovariances (after Bennett, Chua and Leslie, 1997).
60N 1
2
30N
153
46
0 30S 60S 0
60E
120E
180
120W
60W
180
120W
60W
60N 1
30N
2 4
0 30S 60S 0
60E
120E
.25
0
0.25
1
40N
.00
0.75
0 1.0
40N
1.50
3.00 4.50 6.00
0.25
30N
0.7
30N
7.50
0.25
5
1. 50
0
20N
4.5
0 3.0
5 0.2
20N
1.00
0.75
10N
10N
0
0
10S
10S 110E
130E
150E
170E
110E
130E
150E
170E
Figure 5.4.3 As for Fig. 5.4.2; closeups over the S. China Sea (after Bennett, Chua and Leslie, 1997).
about the ﬁfth and ﬁnal “outer” iterate of the inverse estimate. Upper level zonal winds for two representers are shown in Fig. 5.4.2, and in closeup in Fig. 5.4.3. Their striking anisotropy is a consequence of shearing by supertyphoon winds. The eigenvalues of the M × M representer matrix R and its stabilized form P = R + C , where C is
154
5. The ocean and the atmosphere
200 hPa
(T=24)
300 hPa
40N
40N
40N
30N
30N
30N
20N
20N
20N
10N
10N
10N
0
0
0
10S
10S
10S
110E
130E
150E
170E
110E
200 hPa
(T=18)
150E
170E
110E
300 hPa
40N
40N
30N
30N
20N
20N
20N
10N
10N
10N
0
0
0
10S
10S
10S
130E
150E
170E
110E
130E
150E
170E
110E
40N
40N
30N
30N
30N
20N
20N
20N
10N
10N
10N
0
0
0
10S
10S
10S
130E
150E
170E
110E
200 hPa
130E
150E
170E
110E
300 hPa
40N
40N
30N
30N
30N
20N
20N
20N
10N
10N
10N
0
0
0
10S
10S
10S
130E
150E
170E
110E
130E
170E
130E
150E
170E
130E
150E
170E
150E
170E
850 hPa
40N
110E
150E
850 hPa
300 hPa
40N
110E
130E 850 hPa
30N
200 hPa
(T=0)
130E
40N
110E
(T=12)
850 hPa
150E
170E
110E
130E
Figure 5.4.4a The ﬁrst orthonormalized eigenvector z1 of the symmetric positivedeﬁnite matrix P = R + C .
the data error covariance matrix, may be seen in BCL1. Recall that R was calculated on the relatively coarse grid as a preconditioner for indirect inversion on the ﬁne grid. Given the assumed levels of data error, there are only about 200 effective degrees of freedom in the observations. Thus only about 200 iterations would be needed in order to solve (3.1.6). The coarse grid precondition reduced the number to below 15. The ﬁrst and fourth leading normalized eigenvectors of P are shown in Fig. 5.4.4. They are associated with the ﬁrst and fourth largest eigenvalue of P: see §2.5 and
5.4 Tropical cyclones (2)
200 hPa
(T=24)
300 hPa
40N
40N
30N
30N
30N
20N
20N
20N
10N
10N
10N
0
0
0
10S
10S
10S
130E
150E
170E
110E
200 hPa
(T=18)
150E
170E
110E
300 hPa
40N
40N
30N
30N
20N
20N
20N
10N
10N
10N
0
0
0
10S
10S
10S
130E
150E
170E
110E
130E
150E
170E
110E
300 hPa
40N
40N
30N
30N
30N
20N
20N
20N
10N
10N
10N
0
0
0
10S
10S
10S
130E
150E
170E
110E
200 hPa
130E
150E
170E
110E
300 hPa
40N
40N
30N
30N
30N
20N
20N
20N
10N
10N
10N
0
0
0
10S
10S
10S
130E
150E
170E
110E
130E
170E
130E
150E
170E
130E
150E
170E
150E
170E
850 hPa
40N
110E
150E
850 hPa
40N
110E
130E 850 hPa
30N
200 hPa
(T=0)
130E
40N
110E
(T=12)
850 hPa
40N
110E
155
150E
170E
110E
130E
Figure 5.4.4b The fourth eigenvector z4 .
especially (2.5.4). These are the ﬁrst and fourth most stably observed wind patterns. There is crosscorrelation between wind components u and v; there is autocorrelation in time and there is autocorrelation in height. The amplitudes of the winds are asymmetric with respect to the centers of the tropical cyclones, evidently as a consequence of strongly asymmetrical advection. The two eigenvectors display markedly different ﬂow topologies. Variational methods are capable of extracting nonintuitive covariance structure from dynamics, even if the use of such methods for actual assimilation or analysis cannot be afforded in real time. The realtime imperative is most demanding
156
5. The ocean and the atmosphere
in numerical weather prediction, but is far less demanding in seasonaltointerannual climate prediction.
5.5
ENSO: testing intermediate coupled models
The Tropical Atmosphere–Ocean array developed for the Tropical Ocean–Global Atmosphere experiment, or TOGATAO array, is providing an unprecedented in situ data stream for realtime monitoring of tropical Paciﬁc winds, sea surface temperature, thermocline depths and upper ocean currents. For a tour of the project, and for data display and distribution, see www.pmel.noaa.gov/tao/home.html. The data are of sufﬁcient accuracy and resolution to allow for a coherent description of the basin scale evolution of these key oceanographic variables. They are critical for improved detection, understanding and prediction of seasonal to interannual climate variations originating in the Tropics, most notably those related to the El Ni˜no Southern Oscillation (ENSO) (McPhaden, 1993, 1999a,b). The freelydistributed TAO display software provides gridded SST and 20◦ isotherm depth (Z20) using an objective analysis procedure. The ﬁrstguess ﬁelds are those of Reynolds and Smith (1995) for SST; a combination of Kessler (1990) expendable bathythermograph (XBT) analyses and Kessler and McCreary (1993) conductivity, temperature, and depth analyses for Z20, and Comprehensive Ocean–Atmosphere Data Set analyses (Woodruff et al., 1987) for surface winds. The procedure is univariate and involves bilinear interpolation followed by smoothing with a gappy running mean ﬁlter (Soreide et al., 1996). Given this splendid and growing data set (see the TOGATAO website), the question arises: can it be better analyzed by generalized inverse methods? That is, can it be better interpolated, or more generally smoothed using a dynamical model as a guide? The question is addressed by Kleeman et al. (1995) who vary the initial conditions and parameters of an “intermediate” coupled model. Miller et al. (1995) apply the Kalman ﬁlter to a linear intermediate ocean model expanded in its natural Rossby wave modes; dynamical errors or “system noise” are admitted and these are assumed to be uncorrelated in time or “white”. Bennett et al. (1998, 2000, hereafter BI, BII) seek upper ocean ﬁelds and lower atmosphere ﬁelds that provide weighted, leastsquares bestﬁts to 12 and 18 month segments of monthly mean TAO data, and to a nonlinear intermediate coupled model after that of Zebiak and Cane (1987). The model structure is indicated schematically in Fig. 5.5.1; the equations of motion may be found in the references. The dynamical variables are anomalies of current, wind temperature and layer thickness, relative to their respective annual cycles. The oceanic and atmospheric dynamics are linear, save for the presence of anomalous advection of anomalous heat in the oceanic upper layer, for the quadratic dependence of anomalous surface stress upon anomalous wind, and for the parameterization of turbulent vertical mixing in the ocean in terms of a
5.5 ENSO: testing intermediate coupled models
p a /ρag
c a2 /g
ua
157
Figure 5.5.1 A reduced gravity, twoandonehalf layer ocean model coupled to a reduced gravity, oneandonehalf layer atmospheric model (after Bennett et al., 1998).
H (1)+θ(1)
u (1)
T
H (1)
H (2)
H (2)+θ(2)
u (2)
piecewise differentiable switching function. The ocean model domain is a rectangle on the equatorial beta plane: (123.7◦ E, 84.5◦ W) × (29◦ S, 29◦ N). The atmospheric model domain is an entire equatorial zone (29◦ S, 29◦ N). The inclusion of local accelerations in all the momentum equations permits the satisfaction of rigid meridional boundary conditions in the ocean, and the satisfaction of rigid zonal boundary conditions in both the ocean and atmosphere. The inclusion of pseudoviscous stresses permits the satisfaction of no slip and free slip at meridional and zonal boundaries respectively. The generalized inverse of this intermediate coupled model and the TAO data is, again, the weighted leastsquares best ﬁt to the dynamics, the initial conditions and the data. The weights are, as usual, the operatorinverses of the covariances of the dynamical, initial and observational errors. The three error types are assumed mutually uncorrelated. The root mean square data errors are: 0.3◦ for Sea Surface Temperature (SST), 3 m for the 20◦ isotherm depth (Z20) and 0.5 m s−1 for each wind component (u a , v a ). The initial errors are assigned the covariance parameters of the ENSO anomalies themselves (see Kessler et al., 1996), and are assumed mutually uncorrelated. Most difﬁcult of all is the prescription of dynamical error covariances. There will inevitably be errors in the parameterizations of turbulent mixing and exchange processes. In the
158
5. The ocean and the atmosphere
case of intermediate models, there are also errors arising from the neglect of numerically resolvable pseudolaminar processes, such as anomalous advection of anomalous momentum and anomalous layer thickness. The prior dynamical covariances in BI and BII are based solely on the latter type of error, which are readily assessable since these would have the scales of the ENSO anomalies themselves. The scales are taken from Kessler et al. (1996). The functional forms of the covariances are chosen, in the absence of real knowledge, for maximal simplicity and computational efﬁciency. The variances are stationary, zonally uniform and concentrated in the equatorial waveguide. The spatial correlations are bellshaped but anisotropic, while the temporal correlations are Markovian (see §3.1.6). The TAO data are selected from three episodes: April 1994–March 1995 (“Year 1”) covering an anomalously warm (+2.5◦ C) western/central Paciﬁc; April 1995–March 1996 (“Year 2”) covering a mild (−1◦ C) La Ni˜na event, and December 1996–May 1998 (“Year 3”), covering one of the major El Ni˜no events of modern times with an anomalously warm (+5◦ C) eastern Paciﬁc. The inverse solutions ﬁt all the TAO data to within about one standard error. The worst ﬁts occur during the mild La Ni˜na event of Year 2; the best occur during the major El Ni˜no event of Year 3: see Fig. 5.5.2. The inverse circulation ﬁelds are discussed in detail in BI and BII; only the residuals and diagnostics will be reviewed here. Consider for example the dynamical residual r T for the SST equation, shown in Fig. 5.5.3 for September 30, 1994. The quantity plotted is the equivalent surface heat ﬂux ρ1 C p H r T , where ρ1 is the density of sea water, C p its heat capacity, and H the thickness of the ocean surface layer. The contour interval is 20 W m−2 . The prior estimate of 50 W m−2 is very signiﬁcantly exceeded over large regions, mostly on the equator. The zonal scale of 30◦ is that of the corresponding covariance. This ﬁeld of residuals is one day’s distribution of heat sources and sinks that must be admitted in the model if the local rate of change of SST is to be consistent with the TAO data. There are two candidates for r T : the unresolved advective heat ﬂuxes (both horizontal and vertical), and the missing heat exchange between the model ocean and the model atmosphere. The atmospheric component of the coupled model exchanges heat with the oceanic component at the rate Q˙ S , but not vice versa. Radiative feedback from clouds is thereby excluded. The atmospheric budget for geopotential anomaly φ is of the form ∂φ · · · = − Q˙ S = −K T, ∂t
(5.5.1)
where T is the SST anomaly and K is a positive constant. Thus a positive SST anomaly (and therefore atmospheric heating) leads to a decrease in geopotential anomaly. The region of signiﬁcant and positive r T on Sept. 30, 1994 coincides with a positive anomaly T (see BI, Fig. 5). Hence both the model ocean and atmosphere gain heat locally on that day. It must be concluded that r T represents mostly an unresolved convergence
(a)
o
o
5S
4
SST
o
0
6
5N
6
4
4
3
2
2
0
1
2 0 2
D J F M A M J J A S O N D J F M AM
(b)
2
0
4
1
D J F M A M J J A S O N D J F M AM
o
o
5S
o
0
80 60
D J F M A M J J A S O N D J F M AM
5N
150
40
100
20
50
0
0
20
40
Z20
20 0 20
D J F M A M J J A S O N D J F M AM
(c)
ua
50
D J F M A M J J A S O N D J F M AM
o
5S
0
40
o
o
5N
4
4
4
2
2
2
0
0
0
2
2
2
4
D J F M A M J J A S O N D J F M AM
(d)
D J F M A M J J A S O N D J F M AM
o
0
5S
2
1
va 0
1
4
D J F M A M J J A S O N D J F M AM
4
o
D J F M A M J J A S O N D J F M AM o
5N
2
2
0
0
2
2
4
4
6
D J F M A M J J A S O N D J F M AM
D J F M A M J J A S O N D J F M AM
6
D J F M A M J J A S O N D J F M AM
Figure 5.5.2 Time series of the inverse estimate of the anomalous state at three TAO moorings (5◦ S, 0◦ , 5◦ N) along 95◦ W. The centered symbols are the 30day average TAO data. All data of the same type are assigned the same standard error so only one bar is shown per panel, but note that the amplitude scale and bar length vary from panel to panel. Results here are for “Year 3” (December 1996–May 1998): (a) SST, ±0.3 K; (b) Z20, ±3 m; (c) u a , ±0.5 m s−1 ; (d) v a , ±0.5 m s−1 (after Bennett et al., 2000).
30o N 15o N
20 0
0
20
60 20 0
60 100
20 60
20
15o S 0
30o S 140o E 160o E
180
160o W 140o W 120o W 100o W
Figure 5.5.3 Scaled residual for the temperature equation for day 180, Year 1 (September 30, 1994). Contour interval: 20 W m−2 (after Bennett et al., 2000).
160
5. The ocean and the atmosphere
of oceanic heat ﬂux, rather than a neglected exchange with the atmosphere. On “Feb. 30” 1995, the region of signiﬁcantly negative r T (see BI, Fig. 12) coincides with a negative anomaly T , yielding the same conclusion. On Nov. 30, 1994 negative r T coincides with positive T , possibly representing a loss from the ocean to the atmosphere rather than an oceanic heat ﬂux divergence. No clear evidence of loss from the atmosphere to the ocean is seen in Year 1. In principle, both candidates for r T should be accepted. However, scale analysis shows that an oceanic temperature source of strength (ca2 /g H )ρ a K T (ρ1 C p )−1 , ca being the atmospheric phase speed, g being gravity and ρ a the air density, is an order of magnitude smaller than the prior standard deviation for the residual r T . Thus, only oceanic heat ﬂux convergence is a credible candidate for r T . The convergence may be vertical or horizontal. The model’s simple parameterization of heat ﬂux using a simple mixing function is almost certainly signiﬁcantly in error. The linear momentum equations in the model ocean and atmosphere do not support eddies or instabilities such as tropical instability waves that could produce horizontal eddy heat ﬂuxes. It should, however, be pointed out that such waves tend to be weakest during El Ni˜no events (and 1994–1995 is no exception), and that they tend to be strongest east of 150◦ W. Yet Fig. 5.5.3 shows that the maximum SST dynamical residuals r T are near 160◦ W. Nevertheless, the oceanic momentum equations should include horizontal advection, in addition to wellresolved vertical advection and betterparameterized vertical mixing. It is simple to recompute the inverse with the dynamics imposed as strong constraints: the dynamical error variances are set to zero and the iterated indirect representer algorithm is rerun. There are sufﬁciently many degrees of freedom in the initial residuals to enable the inverse to ﬁt the data at some moorings for three months, but nowhere for longer times: see Fig. 5.5.4 (Year 1). Monte Carlo methods may be used to approximate the posterior error covariances: see §3.2. These are relatively smooth and need not be computed on as ﬁne a grid as is used for the inverse itself. A small number of samples should be adequate for such low moments of error, if not for Monte Carlo approximation of the inverse itself. Recall that the representers are themselves covariances (see §2.2.3), and so may be approximated by Monte Carlo methods. Comparisons with representers and inverses calculated with the Euler–Lagrange equations demonstrate the accuracy of sampling methods. Shown in Fig. 5.5.5 are four calculations of SST for Nov. 1994. Daily values are calculated as described below, and then averaged for 30 days. The ﬁrst panel shows the solution of the Euler–Lagrange equations. This is a true ensemble estimate since it is a solution to what are, in effect, the moment equations for the randomly forced coupled model. The second, third and fourth panels are Monte Carlo estimates based on respectively 100, 500 and 1500 samples. It is disturbing that the +2◦ warm pool on the Dateline, characterizing the moderate El Ni˜no of Year 1, is only clearly expressed with 1500 samples. These calculations, variational and Monte Carlo, are all made on the same spatial grid and at the same temporal resolution.
5.5 ENSO: testing intermediate coupled models
o
o
o
SST (0 , 155 W) 2.0
161
o
SST (2 N, 95 W) 2.0 1.0
1.0
0.0 1.0
0.0
2.0 1.0
A M J J A S O N D J F M
o
3.0
A M J J A S O N D J F M o
o
o
Z20 (8 S, 155 W)
Z20 (0 , 155 W) 40.0
20.0
30.0
0.0
20.0 20.0 10.0 40.0 60.0
0.0 A M J J A S O N D J F M
a
o
10.0
o
a
u (8 S, 155 W) 1.0
2.0
0.0
1.0
1.0
0.0
2.0
A M J J A S O N D J F M
o
o
v (8 S, 155 W)
3.0
1.0
A M J J A S O N D J F M
3.0
A M J J A S O N D J F M
Figure 5.5.4 TAO 30day averaged data (centered symbols) and time series of SST, Z20, u a and v a at selected moorings. Solid lines: weak constraint inversion; dashed lines: strong constraint inversion (Year 1, after Bennett et al., 1998).
Monte Carlo approximation of the error covariances indicates a more relaxed state of affairs. Shown in Fig. 5.5.6 are the prior and posterior variances of initial SST errors as functions of longitude and latitude. The great difference between the prior and posterior (or “explained”) variances, with 140 samples, greatly exceeds the sampling error in the prior variance. The small posterior variance implies that the initial SST estimate is reliable. Similar implications hold for the inverse estimates of SST throughout Year 1: see BI, Figs. 17 and 18 for prior and posterior variances for that variable and other coupled model variables. There is, however, a caveat. All these priors and posteriors
(a) 30o N 0.0
0.0
15o N
0.0 1.0
1.0
2.0
0
1.0
1.0
1.0
15o S
0.0
0.0
1.0
30o S 140o E 160o E
180
160o W 140o W 120o W 100o W
(b) 30o N
0.0 0.0
0 .0 0.0
15o N
0.0 0.0
0.0
1.0
1.0
0
0.0 2.0
1.0
0.0
1.0
1.0
15o S 0.0
0.0
30o S 140o E 160o E
180
160o W 140o W 120o W 100o W
(c) 30o N 15o N
0.0
0.0
1.0
0.0
0.0
0.0
0.0
0.0 1.0
2.0
1.0
0
1. 0.00 1.0
1.0
0.0
30 S
0.
1.0
1.0
0.0 o
0.0
1.0
15o S 0
140o E 160o E
180
160o W 140o W 120o W 100o W
(d) 30o N
0.0 0.0
0.0
15o N
0.0
0
1.0 1.0
2.0
1.0 2.01.0
1.0
15o S 0.0
0.0
30o S
1.0
1.0
140o E 160o E
180
160o W 140o W 120o W 100o W
Figure 5.5.5 Inverse estimate of daily SST averaged for month 8, Year 1 (Nov. 1994); (a) as a solution of the Euler–Lagrange equations, (b) as a statistical simulation with 100 samples, (c) 500 samples, (d) 1500 samples.
5.5 ENSO: testing intermediate coupled models
5
(a) Prior (Hyp.) Prior (2025) Prior (140) Posterior (140)
3
2
(K )
4
2 1 0
4
140E
180
140W
100W
(b)
2
(K )
3
Prior (2025) Prior (140) Posterior (140)
2
163
Figure 5.5.6 Statistical simulations of (a) equatorial and (b) meridional proﬁles of prior and posterior error variances for initial SST. The level broken line in (a) is the hypothesized initial equatorial error variance of 4K2 . In (b), the hypothesis is indistinguishable from the solid line. The numbers in parentheses indicate the number of samples (after Bennett et al., 1998).
1 0 30S
15S
0
15N
30N
are based on the null hypothesis for the prior errors in the initial conditions, data and dynamics. Thus the posteriors derived from the hypothesized priors cannot be trusted until the null hypothesis has survived signiﬁcance tests. The prior and posterior values of the penalty functional are test statistics for the null hypothesis: see §2.3.3. Their values for the three El Ni˜no–La Ni˜na episodes are shown in Table 5.5.1, which is taken from BII. They are calculated using, where needed, a Monte Carlo estimate of the full representer matrix R. Note that the prior J F and posterior Jˆ need only the prior data misﬁt vector h, the speciﬁed measurement error covariance matrix C , and the representer coefﬁcient vector βˆ which may be obtained without explicit construction of R: see §3.1.4. With the exception of Jˆ , the expectations and variances of these statistics all do depend explicitly upon R. That the actual values of J F in all three “Years” are signiﬁcantly less than their expected values, suggests that the forward model is far more accurate than hypothesized. Inversion would seem unnecessary. Yet, the actual values of Jˆ for the three years exceed their expected values by 15, 16 and 49 standard deviations, respectively. On the other hand, rescaling the standard deviations of the errors in the null hypothesis by 1.40, 1.44 and 2.08, respectively, would yield values of Jˆ equal in each case to the expected value M given by the number of data. Such rescalings could hardly be contested, in light of the uncertainties involved in developing the null hypothesis. A fourth year of data, for another El Ni˜no event, is needed in order to obtain at least one independent test of the last upward rescaling. It would serve little purpose, as the dynamical residuals already dominate the term balances: see Fig. 5.5.7. A rescaling of the priors might well yield a statistically selfconsistent analysis of TAO data using an intermediate coupled model, but the model constraint would be so “slack” that it would provide no dynamical insight into ENSO. Fully stratiﬁed models are needed, with ﬁne vertical resolution and good estimates of moments of errors in the turbulence parameterizations. The calculations described above involve about 4 × 107 control variables or residuals; there are about 2500 monthlymean data in the 12month episodes (Years 1 and 2)
5. The ocean and the atmosphere
164
Table 5.5.1 (a) Expected and actual values of components of the reduced penalty functional for the intermediate coupled model (those values in parentheses are numbers of data rather than expected values of data penalties); (b) standard deviations (after Bennett et al., 2000).
JF Jˆmod JˆSST Jˆu a Jˆva JˆZ 20 Jˆdata Jˆ
(a) Expected and actual values Year√1 Year√2 M = 2644, 2M = 73 M = 2624, 2M = 72
Year√3 M = 4008, 2M = 89
Expected 16 0000 1015 (689) (624) (624) (687) 1609 2624
Expected 246 132 1458 (1088) (931) (931) (1058) 2550 4008
Actual 35 708 1614 419 453 396 820 2088 3702
Year 1 JF Jˆmod Jˆdata Jˆ
Expected 15 6000 1022 (700) (628) (628) (680) 1622 2644
Actual 32 028 1650 717 383 380 678 2160 3810
(b) Standard deviations Year 2
136 000 38 60 72
200 (a)
131 000 38 59 73
rT Tt
100
(1)
u (T+T)x (1)
0
u Tx MTz
100 200
60
σ +σ D J FMAM J J A SOND J FMAM
(b)
rθ(2) θ
40
(2)
t (2)
(2)
εθ
0
σ +σ
20 40
(2)
H (u
20
D J FMAM J J A SOND J FMAM
(2)
x+v y)
Actual 132 140 3680 995 1546 940 1129 4614 8322
Year 3 146 057 72 45 89
Figure 5.5.7 Time series of Year 3 term balances for the intermediate coupled model for (a) the anomalous SST equation in W m−2 at (135.85◦ W, 3.5◦ S) and (b) for the anomalous lowerlayer thickness equation in 10−6 m s−1 at (156.38◦ W, 0.5◦ S). All are daily values, spaced thirty days apart and joined by line segments for clarity. The standard error σ in (a) is 54 W m−2 ; in (b) it is 8 × 10−6 m s−1 (after Bennett et al., 2000).
5.6 Sampler of oceanic and atmospheric data assimilation
SST
165
Z20
3
20 10
2
m
K
0 1
10 20
0 30 1
0
90
180
270
40
360
0
90
day
180
270
360
270
360
day
a
u
v
a
4
10
2
ms
ms
1
1
5 0
0 2 5
0
90
180
270
360
4
0
90
day
180
day
Figure 5.5.8 Time series of SST, Z20 and u a , v a at (180◦ ,0◦ ) for Year 1. Diamonds: 30day averaged data; solid lines: corresponding inverse estimate; triangles: ﬁveday averaged data; broken lines: corresponding inverse estimate.
and 4000 in the 18month episode (“Year” 3). The inversion for Year 1 has been repeated, using ﬁveday averaged data in place of 30day averages. There are accordingly about 15 000 of the former. A time series of the inverse SST is shown in Fig. 5.5.8. The inverse is unable to ﬁt the data to within the 0.3◦ standard error of measurement, not because the data are of lower quality but because the time scales of the null hypothesis and intermediatemodel dynamics are too long. The ﬁgure emphasizes that the inverse is indeed a “ﬁxed interval smoother”, and that the number M of data is not a serious restriction on the indirect representer method. Contemporary matrix manipulation techniques are not severely strained at M = 105 (see e.g., Egbert, 1997, §2.3; Daley and Barker, 2000).
5.6
Sampler of oceanic and atmospheric data assimilation
5.6.1
3DVAR for NWP and ocean climate models
Operational Numerical Weather Prediction relies upon timely, robust and accurate estimates of initial conditions. For example, the US Navy’s Fleet Numerical Meteorology
166
5. The ocean and the atmosphere
and Oceanography Center (www.fnmoc.navy.mil) produces a global sixday forecast every 12 hours using the global spectral model NOGAPS (Hogan and Rosmond, 1991), and many regional threeday forecasts every 12 hours using the regional ﬁnitedifference model COAMPS (Hodur, 1997). The initial ﬁelds or “analyses” are created by interpolating vast quantities of atmospheric data collected by the US Navy, and also by civilian meteorological agencies around the world. An elementary introduction to “Optimal Interpolation” may be found in §2.2.4. The operational “OI” scheme at FNMOC involves multiple random ﬁelds; the priors or ﬁrst guesses or “backgrounds” for these ﬁelds are previous predictions for that time, while data are admitted through a time window of minus and plus three hours (implying that the analysis time is at least three hours in the past). For comprehensive details of FNMOC’s MultiVariate Optimal Interpolation analysis (“MVOI”), see Goerss and Phoebus (1993). Most importantly, the prior covariances for the velocity and geopotential ﬁelds assume geostrophy (5.1.39). However, the resulting analysis is nongeostrophic since the background ﬁeld is a prediction made by a Primitive Equation model. The essential computational task in an OI scheme is solving the linear system {Cq + C }β = d − u F
(5.6.1)
for the coefﬁcients β of the covariances Cq (x, t) in (2.2.22). Recall that d is the vector of data, while u F is the vector of “measured” values of the background u F (x, t). Note that MVOI replaces the single spatial coordinate x in (2.2.22) with three spatial coordinates, one of which may be pressure as in (5.1.1). The dimension of the system is the number of data, which can be in excess of 105 in a global analysis. MVOI reduces the dimension by analyzing the data in regions. The coefﬁcient matrix in (5.6.1) is symmetric and positive deﬁnite, so MVOI solves the system by Cholesky factorization (Press et al., 1986, §2.9). Should small negative eigenvalues be encountered, MVOI arbitrarily increases the diagonal of C , that is, it increases the variances of the measurement errors until positivity is restored. The solution of (5.6.1) also mimizes the penalty function J [β] =
1 T β {Cq + C }β − β T (d − u F ). 2
(5.6.2)
Preconditioned conjugate gradient searches (Press et al., 1986, §10.6) for the unique minimum of J may be efﬁciently implemented on multiprocessor computers. System dimensions of 105 or larger can be managed, and so regionalization is not necessary. This algorithm for implementing OI has become known as “3DVAR”, and is being introduced at many NWP centers such as FNMOC (Daley and Baker, 2000), at the United Kingdom Meteorological Ofﬁce (Lorenc et al., 2000) and at the NASA Global Modeling and Assimilation Ofﬁce (www.polar.gsfc.nasa.gov). Oceanographers can learn much from these operational NWP centers, concerning the realtime qualitycontrol of vast data sets and the devising of prior multivariate covariances. A fast algorithm for the spatial “convolution” (3.1.25), essential to efﬁcient implementation of physically realizable inverse models, is also required for 3DVAR.
5.6 Sampler of oceanic and atmospheric data assimilation
167
Passi et al. (1996) propose a productpolynomial algorithm for convolving with the “bellshaped” covariance (3.1.24), as an alternative to solving the pseudoheat equation (3.1.26) subject to (3.1.27). The need to initialize and validate climate models is stimulating major applications of “OI” to the global upper ocean. Carton et al. (2000a, 2000b) apply 3DOI to 45 years of upperocean data. Their backgrounds for the ﬁelds of temperature, salinity and current are the predictions of the standard Primitive Equation model MOM2 from the NOAA Geophysical Fluid Dynamics Laboratory (www.gfdl.gov/~smg/ MOM/MOM.html). Carton et al. pay careful attention to the mean background errors or “biases” following Dee and da Silva (1998), and also hypothesize a detailed but structurally simple multivariate covariance for the background errors. 5.6.2
4DVAR for NWP and ocean climate models
Inverse modeling often involves compromises. A common assumption is that the equations of motion are exactly correct, and that only the initial conditions and some dynamical parameters should be perturbed in order to ﬁt the data. This would seem entirely reasonable if the smoothing or assimilation interval of interest is rather less than the evolution time scale of the dynamics: say, rather less than three days on synoptic scales in the midlatitude troposphere, and rather less than three months on planetary scales in the tropical Paciﬁc Ocean. There are two major variational inverse models of this kind, which are “strong constraint” assimilations in the terminology of Sasaki (1970). The ﬁrst is the “4DVAR” program in support of operational Numerical Weather Prediction at the European Centre for Mediumrange Weather Forecasting (ECMWF). The project is described in a major series of papers: Rabier et al. (2000), Mahfouf and Rabier (2000) and Klinker et al. (2000). The model is a spectral representation of the global atmosphere, with about 107 spatial variables per time step. Variational assimilation is performed in sixhour intervals, from t − 3 hours to t + 3 hours, with vast amounts of tropospheric data being smoothed throughout the interval. There are O(105 ) surface data alone. The perturbed state at time t = 0 hours becomes the initial condition for a forecast out to t = 168 hours, and the skill of the forecast is the basis for assessing the utility of the variational assimilation. The second project is the “Estimation of the Circulation and Climate of the Ocean” (ECCO) Consortium (Stammer et al., 2000). The estimation is based on the MIT nonhydrostatic General Circulation Model (Marshall et al., 1997a,b). The tangent– linear and corresponding adjoint operators are constructed with a symbolic algorithm (Giering and Kaminsky, 1997), as described in Marotske et al. (1999). Ocean circulation is sustained by surface ﬂuxes, both in reality and in models. These ﬂuxes are poorly known, and so it is desirable that they should be perturbed along with the initial ◦ conditions in the search for a better ﬁt to data. With 14 resolution globally, there are about 2 × 108 initial variables, and about 107 surface ﬂuxes per time step. The latter need not be perturbed independently at every time step. Even so, the computational
168
5. The ocean and the atmosphere
challenge is clearly enormous, yet impressive progress is being made with 1◦ grids and years of data (Stammer et al., 2000). 5.6.3
Correlated errors
The penalty functional (1.5.7) for controlling the toy model (1.2.6)–(1.2.8) does not contain products of b(t) and f (x, t), for example. The statistical interpretation (2.2.3) is the hypothesis that these boundary errors and forcing errors are uncorrelated. Bogden (2001) argues that boundary ﬂow errors can be correlated to wind errors inside an ocean region, since both can be correlated to wind errors outside the region. Thus the hypothesis should include a nonvanishing crosscovariance: C f b (x, t, s) = E( f (x, t)b(s)).
(5.6.3)
For consistency, we should change the notation for autocovariances, from (2.2.2) to C f f (x, t, y, s) = E( f (x, t) f (y, s)),
(5.6.4)
for example. The estimator of maximum likelihood for such multivariate normal ﬁelds is % &−1 Cff Cfb • f J [u] = ( f •, b ∗ ) + ···. (5.6.5) ∗ b Cb f Cbb The matrix inverse is deﬁned as a matrixvalued kernel: see (1.5.11) and (1.5.12).
Exercise 5.6.1 Assume the other errors are uncorrelated with f or b or with each other, that is, assume the rest of the estimator (1.5.9) is unaltered. Derive the Euler–Lagrange equations for the penalty functional (5.6.5). Show that the weighted residual still obeys (1.3.1)–(1.3.3), but the inverse estimates for the dynamical and boundary residuals are, respectively: fˆ = C f f • λ + C f b ∗ λ, bˆ = Cb f • λ + Cbb ∗ λ,
(5.6.6) (5.6.7)
where the blob products are evaluated inside the region, while the star products are evaluated on the boundary. 5.6.4
Parameter estimation
The constant phase speed c in the toy model (1.1.1) has been kept ﬁxed up to now. Yet the ﬁt to data may be improved by varying c. To this end, the penalty functional J [u, c] in (1.5.9) may be augmented (Bennett, 1992, §10.2): 2 K[u, c] = σ −2 f (c − c0 ) + J [u, c],
(5.6.8)
5.6 Sampler of oceanic and atmospheric data assimilation
169
where c0 is a prior for c, and σc2 is the hypothesized variance of the prior error. Varying K with respect to c and u(x, t) yields the extremal condition cˆ = c0 + σc2
∂ uˆ •λ ∂x
(5.6.9)
and the Euler–Lagrange equations for J as before. Note that uˆ and λ depend upon cˆ , so (5.6.9) is highly nonlinear even though the dynamics of the toy model are linear. However, iteration schemes for solving (5.6.9) are readily devised (Eknes and Evensen, 1997, who consider linear Ekman layer dynamics with an unknown eddy viscosity), and interlaced with iterated representer algorithms in the case of nonlinear dynamics (Muccino and Bennett, 2002, who consider the Korteveg–DeVries equation with unknown parameters for phase speed, amplitude and dispersion). Convergence of these schemes is in no way assured.
5.6.5
Monte Carlo smoothing and ﬁltering
Consider the toy nonlinear model (3.3.1)–(3.3.3), where the random inputs f , i and b have covariances C f , Ci and Cb respectively. The methods of §3.2.4 may be used to generate pseudorandom samples of the inputs consistent with their respective covariances. A pseudorandom sample of the state u(x, t) is then obtained by integrating (3.3.1)– (3.3.3). Sample estimates of the expectation Eu(x, t) and covariance Cu (x, t, y, s) follow from repeated generation and integration. The sample moments of u may then be used for space–time optimal interpolation of data collected in some time interval 0 < t < T , as outlined in §2.2.4. The prior for the OI or best linear unbiased estimate (2.2.22) would not be u F (x, t), but rather the sample estimate of Eu(x, t), while the covariance Cq (x, t, y, s) would be the sample estimate of Cu (x, t, y, s). As a consequence of the nonlinearity of (3.3.1), the OI estimate is not an extremum of the penalty functional (1.5.9), even if W f were related to C f through (1.5.11), (1.5.12), etc. That is, the OI estimate is not the solution of an inverse model. Nevertheless the attraction of such “Monte Carlo smoothing” is obvious: there is no need to linearize the dynamics, nor is it necessary to derive the adjoint dynamics. Storing Cu (x, t, xm , tm ), where (xm , tm ) is a data point, may not be feasible for all x, t and for 1 < m < M, but it may be feasible to store Cu (x, tm , y, tm ), for all x, y and for one time tm . Data collected at the time tm may thus be optimally interpolated in space, provided it is assumed that the data errors are uncorrelated in time. This Monte Carlo ﬁltering method has become known as the “Ensemble Kalman Filter” or EnKF (Evensen, 1994). For its application to operational forecasting of the North Atlantic Ocean, see http://diadem.nersc.no/project; for application to seasonaltointerannual forecasting of the Tropical Paciﬁc Ocean, see Keppenne (2000). For a careful comparison of the computational efﬁciency of the EnKF with that of the indirect iterated representer algorithm, in the context of an hydrological model and satellite observations of soil moisture, see Reichle et al. (2001, 2002).
170
5. The ocean and the atmosphere
Exercise 5.6.2 Why is it necessary in the EnKF to assume that the errors in data collected at different times are uncorrelated? 5.6.6
Documentation
The most important but most neglected aspect of ocean modeling is documentation. It is no exaggeration to state that if a model has not been documented, it does not exist. Documentation is, of course, a tedious chore and the scarce resources for academic research rarely support anything so pedestrian. One outstanding exception is the Modular Ocean Model version 3, or “MOM3” (www.gfdl.gov/~smg/MOM/MOM.html). Inverse models are vastly more complex (see for example Figs. 3.1.1 and 3.1.2), so their documentation is even more important and even more neglected. Ide et al. (1997) have made the following reasonable proposal of a standard notation for data assimilation. All models are eventually subject to numerical approximation of one form or another, so in computational practice the state x is a vector of ﬁnite dimension I : x = (x1 , x2 , . . . , x I ).
(5.6.10)
The value of I is the product of the number of ﬂuid variables (velocity components, temperature, pressure, etc.) and the number of computational degrees of freedom in space (the number of grid points, in a ﬁnitedifference model). The model evolution in a single time step is ' ( x bn = Mn x bn−1 , (5.6.11) where Mn is in general a nonlinear operator, and the initial vector is x b0 . The subscript n in (5.6.11) indicates the state vector or nonlinear operator at time tn . The superscript b indicates that the vector will be the background for an optimal estimate of the state. Thus x bn , (0 ≤ n ≤ N ) lumps the ﬁeld u F (x, t), (0 < x < L , 0 ≤ t ≤ T ) of §1.1.1. Observations are taken at selected times tn j ; at any such time, these data comprise a vector of dimension K : yoj = y oj1 , y oj2 , . . . , y oj K . (5.6.12) The measurement functional Hk is in general nonlinear; ' ( ybj ≡ H j xbn j ,
(5.6.13)
for example, is the measured value of the background at time tn j . The stage is now set for deﬁning an inverse model in terms of a weighted leastsquares penalty function (Uboldi and Kamachi, 2000), analogous to (1.5.9). The standard notation proposed by Ide et al. (1997) is becoming widely accepted. This aids highlevel dialogue, at the expense of insight into dynamical detail. For example, it is not obvious from the abstract ﬁnitedimensional equation (5.6.11) that
5.6 Sampler of oceanic and atmospheric data assimilation
171
the tangent linearization of the evolution operator Mn may be unphysical: see §3.3.4. Nor is it obvious that spatial and temporal irregularity arising from unsuitable weighting (see §2.6), or from illposedness of the forward model (see Chapter 6), are fundamental issues. Documentation will remain a crucial challenge, even with the advent of standard notations such as that of Ide et al. (1997). Preliminary documentation exists (Chua and Bennett, 2001) for IOM (Inverse Ocean Model), a modular code for the iterated, indirect representer algorithm of Fig 3.1.2. The software engineering for such modular systems must accommodate a wide range of models and modeling practices, and yet retain high computational performance. An elegant graphical representation for the Project d’Assimilation par Logiciel Multim´ethodes (“PALM”), a universal coupler of models and data, has been devised by Lagarde et al. (2001) as an aid to software development.
Chapter 6 Illposed forecasting problems
The “toy” forward model introduced in Chapter 1 deﬁnes a wellposed mixed initial value–boundary value problem. The associated operator (wave operator plus initial operator plus boundary operator) is invertible, or nonsingular. Specifying additional data in the interior of the model domain renders the problem overdetermined. The operator becomes uninvertible, or singular. The difﬁculty may be resolved by constructing the generalized inverse of the operator, in a weighted leastsquares sense. An important class of regional models of the ocean or atmosphere deﬁnes an illposed initialboundary value problem, regardless of the choice of open boundary conditions. All ﬂow variables may as well be speciﬁed on the open boundaries. The excess of information may be regarded as data on a bounding curve, rather than at an interior point. The difﬁculty may again be resolved by constructing the generalized inverse in the weighted leastsquares sense. The Euler–Lagrange equations form a wellposed boundary value problem in space–time. Solving them by forward and backward integrations is precluded, since no partitioning of the variational boundary conditions yields wellposed integrations. The penalty functional must be minimized directly. Things are different if the open region is moving with the ﬂow.
6.1
The theory of Oliger and Sundstr¨om
We have been assuming that our forward model constitutes a wellposed problem. That is, just sufﬁcient information is given about the forcing F(x, t), the boundary values B(t) and initial values I (x) in order to ensure the existence of a solution that is unique
172
6.2 Open boundary conditions
173
for each choice of F, B and I , and which depends continuously upon smooth changes to F, B and I . For linear models, the proof of uniqueness also implies continuous dependence, and even implies existence (Courant and Hilbert, 1962, Ch. VI, §10). Sometimes we are able to establish existence (and the other requirements) by displaying the general solution explicitly. If the explicit solution has been obtained after making an assumption about the solution, such as the variables being separable, then uniqueness must be established ﬁrst. Uniqueness and continuous dependence can be established for some nonlinear models, such as the shallowwater equations (Oliger and Sundstr¨om, 1978), but existence is usually an open question. Existence has been established for inviscid and viscous incompressible ﬂow in the plane (Ladyzhenskaya, 1969). These results have been extended to certain quasigeostrophic ﬂows (Bennett and Kloeden, 1981). In this section we shall brieﬂy review the uniqueness of solutions of some simple linear models, and then use these ideas to determine the number of boundary conditions needed at open boundaries. Following Oliger and Sundstr¨om (1978), it will be shown that there is a difﬁculty with the Primitive Equations, but it will be argued that the illposedness can be resolved by generalized inversion of the openocean Primitive Equation model. In subsequent sections we shall address the special methods for ﬁnding the generalized inverse, given that the forward model is illposed.
6.2
Open boundary conditions for the linear shallowwater equations
In §1.1.2 we veriﬁed that the initialboundary value problem for the onedimensional linear wave equation of §1.1.1 has a unique solution. Now consider a linear shallowwater model: ∂u = −g∇h, ∂t
(6.2.1)
∂h = −H ∇ · u, ∂t
(6.2.2)
where u = u(x, t) is a planar velocity ﬁeld, x is a point in the plane, t is time, h = h(x, t) is the sealevel disturbance, g is the gravitational acceleration, H is the mean depth, and ∇ = ( ∂∂x , ∂∂y ). Suitable initial conditions are u(x, 0) = 0,
(6.2.3)
h(x, 0) = 0.
(6.2.4)
We need not include forcing in (6.2.l) or (6.2.2), nor nonzero initial values in (6.2.3) and (6.2.4), since we are interested in the difference between two solutions having the same forcing, initial values and boundary values.
174
6. Illposed forecasting problems
It follows from (6.2.1) and (6.2.2) that d 1 2 1 2 H  u  + gh da = −g H hu · nˆ ds, dt 2 2 D
(6.2.5)
B
where D is the spatial domain, B is its boundary, da is an area element in D, ds is an arc element on B, and nˆ is an outward normal on B. It is clear from (6.2.3) and (6.2.4) that the area integral on the lhs of (6.2.5) vanishes at t = 0, so if the rhs is nonpositive for all t ≥ 0, then the area integral vanishes for all t ≥ 0; hence u(x, t) = 0 and h(x, t) = 0. Uniqueness would then be established. That is, hu · nˆ ≥ 0
(6.2.6)
on B, for all t ≥ 0, would ensure uniqueness. For example: (i) specify u · nˆ = 0 on B or (ii) specify h = 0 on B or (iii) specify u · nˆ = α

on B,
g h H
where α is a positive constant. Notice that one of (i), (ii) or (iii) would sufﬁce for uniqueness; there is no need to specify both the normal velocity and the sealevel elevation. That would overdetermine the solution.
Exercise 6.2.1 Consider the difference between two solutions of the linear shallowwater equations, corresponding to two different sets of forcing, initial and boundary values. Show that the total energy of the difference is controlled by the differences in the inputs. Hint: let 12  F (t) ≡  F(x, t) 2 da ; D
it may be shown that ≤  F  ·  v  . F · v da D
6.3 Inﬂow and outﬂow
175
Advection: subcritical and supercritical, inﬂow and outﬂow
6.3
In a small step towards nonlinearity, let us now include a constant advecting velocity U in the linear shallowwater equations: ∂u + U · ∇u = −g∇h, ∂t ∂h + U · ∇h = −H ∇ · u . ∂t Hence
(6.3.1) (6.3.2)
d 1 dE ≡ (H u2 + gh 2 ) da dt dt 2 D 1 (H u2 + gh 2 )U · nˆ + g H hu · nˆ ds. =− 2
(6.3.3)
B
In order to establish uniqueness, we must arrange for the rhs of (6.3.3) to be nonpositive. The integrand in (6.3.3) is a quadratic form: I cnˆ u 1 T , (6.3.4) { } = H (u , σ h) 2 cnˆ T U σh 1
1
ˆ I is the 2 × 2 unit matrix and c = (g H ) 2 . We must where σ = (g/H ) 2 , U = U · n, determine if the 3 × 3 matrix in (6.3.4) is deﬁnite or indeﬁnite. Its eigenvalues λ satisfy U − λ 0 cnˆ 1 0 U −λ cnˆ 2 = 0, (6.3.5) cnˆ 1 cnˆ 2 U −λ that is, (U − λ){(U − λ)2 − c2 } = 0. Hence the three eigenvalues are λ0 = U,
λ± = U ± c.
Note that these are deﬁned at each point on the boundary, and that U ≡ U · nˆ is the local value. There are several cases to consider: I. OUTFLOW: U >0 (a) SUPERCRITICAL: U >c>0 All three eigenvalues are positive. So, regardless of the boundary values, dE ≤ 0 and uniqueness follows with dt NO BOUNDARY CONDITIONS.
176
6. Illposed forecasting problems
(b) SUBCRITICAL: 0 0, λ− = U − c < 0. Hence ddtE ≤ 0 provided (uT , σ h) is orthogonal to the eigenvector µ− associated with λ− . This may be assured with ONE BOUNDARY CONDITION. In this case U − λ− = U − (U − c) = c, so µ− = (µ, ν, ξ )T satisﬁes
c 0 cnˆ 1
0 c cnˆ 2
µ cnˆ 1 cnˆ 2 ν = 0. ξ c
(6.3.6)
A nonnormalized choice is µ −nˆ 1 ν = −nˆ 2 , ξ 1
(6.3.7)
−nˆ 1 (u, v, σ h) −nˆ 2 = 0, 1
(6.3.8)
hence we require
or u · nˆ = σ h
(6.3.9)
II. INFLOW: U 0, λ− = U − c < 0: two eigenvalues are negative, so we need TWO BOUNDARY CONDITIONS.
Exercise 6.3.1 Find µ0 , µ− .
6.4 Illposedness
177
The above results may be summarized: SUBCRITICAL SUPERCRITICAL
(U  < c) (c < U )
INFLOW (U < 0) 2 3
OUTFLOW 1 0
(U > 0)
Again, the situation will in general vary around the boundary.
Exercise 6.3.2 Consider the cases U = 0, U  = c.
We conclude that it is possible to determine the correct number of boundary conditions for a linearized shallowwater model. That is, we may determine the number that ensures the uniqueness of solutions, and hence continuous dependence upon the inputs.
6.4
The linearized Primitive Equations in isopycnal coordinates: expansion into internal modes; illposed forward models with open boundaries
The Primitive Equations were presented in §5.1 in isobaric or “pressure” coordinates. Let’s denote these coordinates by (x , y , p, t ). Now consider isopycnal or “density” coordinates (x, y, α, t), where α is the speciﬁc volume or inverse density: α ≡ ρ −1 . Deﬁne the Montgomery potential m by m = αp + φ, where φ = gz is the geopotential.
Exercise 6.4.1 Show that ∂u + u · ∇u = −∇m, ∂t ∂m = p, ∂α ∂2 p ∂p +∇· u = 0. ∂t∂α ∂α
(6.4.1) (6.4.2) (6.4.3)
Explain the meaning of the partial derivatives. Note that there is no “diapycnal” advection in (6.4.1) and (6.4.3). This is a consequence of (1) the combined ﬁrst and second laws of thermodynamics: T dη = de + pdα, where T is temperature, η is entropy, and e is internal energy;
(6.4.4)
178
6. Illposed forecasting problems
(2) the assumption of isentropic motion; (3) the assumption that the internal energy of the ﬂuid is constant. The isopycnal form of the Primitive Equations need not be so restrictive, but this form sufﬁces for our immediate purpose. Linearize about a uniform horizontal velocity U, a staticstate Montgomery potential m = M(α), and a pressure p = P(α), where P = ddαM : ∂u + U · ∇u = −∇m, ∂t ∂m = p, ∂α ∂p dP ∂2 p +U·∇ + ∇ · u = 0. ∂t∂α ∂α dα
(6.4.5) (6.4.6) (6.4.7)
The ﬁelds u, m and p are now perturbations about U, M and P. Separate variables according to u(x, α, t) = u (x, t)A(α),
m(x, α, t) = m (x, t)A(α), p(x, α, t) = p (x, t)
dA (α). dα
(6.4.8) (6.4.9) (6.4.10)
Derive the separated equations ∂u + U · ∇u = −∇m , ∂t
(6.4.11)
∂m + U · ∇m + c2 ∇ · u = 0, ∂t
(6.4.12)
and d2 A − c−2 dα 2
dP dα
A = 0,
where the separation constant c2 has the dimensions of (speed)2 . The unseparated ocean boundary conditions are: p x, α (a) , t = p (a) (x, t),
(6.4.13)
(6.4.14)
where α (a) is an isopycnal surface in contact with the atmosphere, which is at pressure p (a) ; m (b) (x, t) ≡ m x, α (b) , t = α (b) p (b) + φ (b) , (6.4.15) assuming that the ocean bottom at . z = z (b) (x) = φ (b) g
(6.4.16)
6.4 Illposedness
179
is the isopycnal surface α = α (b) . Hence the separated boundary conditions for perturbations are: dA =0 dα
(6.4.17)
dA dα
(6.4.18)
at α = α (a) ; A=α
at α = α (b) . The system (6.4.11), (6.4.12) is the linearized shallow water equations with phase speed c. The system (6.4.13), (6.4.17), and (6.4.18) comprises a regular Sturm– Liouville problem (Stakgold, 1979), with eigenvalues c0 > c1 > · · · > cn > · · · > 0 and eigenmodes A0 (α), A1 (α), . . . , An (α), . . . .
Exercise 6.4.2 Show that A
(b)
2
b = α(b)
dA dα
2
+ c−2
dP 2 A dα
dα,
(6.4.19)
a
where A(b) = A(α (b) ). Note that ddαP < 0. Hence the external mode, for which A is approximately independent of α, has phase speed c0 satisfying c02 ∼ = α (b) P (b) − P (a) .
(6.4.20)
The internal modes have lower phase speeds: (α) − α (b) 2 ∼ 2 α Jn−2 , cn = c0 1 (a) (b) α + α 2
(6.4.21)
where Jn = 0(n) as n → ∞. For the Southern Ocean, c0 ∼ = 220 m s−1 , c1 = 1 m s−1 , c2 = 0.5 m s−1 , c3 ∼ = 0.3 m s−1 , c4 = 0.2 m s−1 . (6.4.22) In the Antarctic Circumpolar Current, u 0.6 m s−1 .
(6.4.23)
c0 > c1 > u > c2 > c3 . . . .
(6.4.24)
Hence
180
6. Illposed forecasting problems
Consider the two lowest modes: n = 0, 1. The amplitudes un (x, t) , m n (x, t) satisfy the linearized shallow water equations (6.4.11), (6.4.12). The ﬂow is everywhere subcritical so two boundary conditions are needed at inﬂow, and one at outﬂow. For all other modes (n = 2, 3, . . .), the ﬂow is in general supercritical (wherever ˆ > c2 ), so three boundary conditions are needed at inﬂow, while none is needed U · n at outﬂow. The problem is that we don’t usually integrate the Primitive Equations mode by mode; we usually specify boundary conditions at each level, or each value of α. Suppose we choose two BCs at inﬂow and one BC at outﬂow. Then modes 2, 3, . . . are underspeciﬁed at inﬂow and overspeciﬁed at outﬂow. Underspeciﬁcation is often incorrectly eliminated by the imposition of what we intended to be computational boundary conditions; in these circumstances they acquire a dynamical role. Overspeciﬁcation leads to computational noise that is usually suppressed by smoothing the ﬁelds. The illposedness of the open boundary problem for the Primitive Equations cannot be solved, but it can be resolved by generalized inversion.
6.5
Resolving the illposedness by generalized inversion
To demonstrate the approach, it sufﬁces to consider the shallowwater equations (Bennett, 1992; Bennett & Chua, 1994). Prescribe three boundary conditions on the open boundary. This overdetermines the problem, so do not seek an exact solution of the equations of motion. Rather, seek a weighted, leastsquares bestﬁt to all the information. The dynamics are thus ∂u + U · ∇u = −c∇q + µ, ∂t
(6.5.1)
∂q + U · ∇q = −c∇ · u + χ , ∂t
(6.5.2)
where q = m/c, while µ and χ are misﬁts or residuals. A simple penalty functional is
T
J [u, q] = WD
dt (µ2 + χ 2 )
da D
0
+ WB
T dt {u − uB 2 + (q − qB )2 }
ds B
+ WI
0
da {u − uI 2 + (q − qI )2 }
(6.5.3)
D
+ (data penalties), where D is the domain, B is the entirely open boundary, (uB , qB ) are the boundary values, and (uI , qI ) are the initial values. The weighting is simple in the extreme, but clarity is
6.5 Generalized inversion
181
the issue here. A more realistic “nondiagonal” weighting should be used in practice. Note that no boundary conditions are needed at a supercritical outﬂow boundary, so the boundary values there are like data prescribed continuously on curves.
Exercise 6.5.1 Show that the Euler–Lagrange equations for local extrema of J are: ∂µ − U · ∇µ − c∇χ + · · · = 0, ∂t ∂χ − U · ∇χ − c∇ · µ + · · · = 0, − ∂t µ = 0, χ = 0 −
(6.5.4) (6.5.5) (6.5.6)
at t = T , −WD µ + WI (u − uI ) = 0,
(6.5.7)
−WD χ + WI (q − qI ) = 0,
(6.5.8)
ˆ + WB (u − uB ) + WB cnχ ˆ = 0, WD U · nµ
(6.5.9)
ˆ + WB (q − qB ) + WB cnˆ · µ = 0, WD U · nχ
(6.5.10)
both at t = 0, and
both on B. The ellipsis (· · ·) in (6.5.4) and (6.5.5) denotes data terms. The system (6.5.1), (6.5.2), (6.5.4)–(6.5.10) is a boundary value problem in the space–time domain D × [0, T ].
Exercise 6.5.2 Show that if (u, q) and (µ, χ) satisfy the Euler–Lagrange equations (6.5.4)–(6.5.10), then
T dt (µ + χ ) + WB
da
WD D
2
2
dt (u2 + q 2 )
ds B
0
T 0
da (u + q )t=0 + (· · ·) = 0,
+ WI
2
2
2
(6.5.11)
D
where (· · ·)2 denotes a nonnegative contribution from the data sites. It follows immediately from (6.5.11) that µ ≡ 0,
χ ≡0
(6.5.12)
u = 0,
q=0
(6.5.13)
in D × [0, T ],
182
6. Illposed forecasting problems
in B × [0, T ], and u = 0,
q=0
(6.5.14)
in D × {0}. The “forward problem” for (u, q) is (6.5.1), (6.5.2), which together with (6.5.12)–(6.5.14) is overdetermined, but the solution is without question u ≡ 0,
q = 0.
(6.5.15)
It may be concluded that the Euler–Lagrange equations form a wellposed boundary problem in D × [0, T ], even though the original forward model is illposed. The challenge is to ﬁnd a solution algorithm for the inverse, when the forward model is illposed. Backward and forward integrations must be avoided.
Exercise 6.5.3 It may be difﬁcult to accept that generalized inversion can be wellposed, even though the forward model is illposed. Consider a simple example, deﬁned by an ordinary differential equation: dx =1 dt
(6.5.16)
for 0 ≤ t ≤ 1, subject to x(0) = 0,
x(1) = 2.
(6.5.17)
This overdetermined problem does not have a continuous solution. Now seek a leastsquares bestﬁt to (6.5.16) and (6.5.17), with a penalty functional 1 J [x] =
2
dx −1 dt
dt + x(0)2 + (x(1) − 2)2 .
(6.5.18)
0
Derive the Euler–Lagrange equations for the best ﬁt, and verify that they have the unique, continuous solution xˆ (t) =
6.6
1 + 4t . 3
(6.5.19)
State space optimization
Bennett and Chua (1994) used simulated annealing and HMC to resolve the illposedness of an idealized, regional shallowwater model. The illposedness arose from specifying too much data at the open boundaries, thereby mimicking the situation that is inevitable for regional Primitive Equation models. As argued in §6.5, the illposedness
6.7 Wellposedness in comoving domains
183
is resolved by reformulation as a generalized inverse problem, since the associated Euler–Lagrange (EL) equations form a wellposed boundary value problem. The EL equations are efﬁciently solved using representer methods, but these require backward and forward integrations. When the dynamics are shallowwater, it is always possible to partition the EL boundary conditions so that both the backward integration and the forward integration are wellposed. Such a partitioning is not possible for Primitive Equation dynamics. So, in the interest of developing generalized inversion techniques alternative to solving the EL equations, Bennett and Chua (1994) minimized the nonlinear shallowwater penalty functional by direct attack with simulated annealing and HMC. Numerical experiments with synthetic data supported the argument that the inverse is wellposed. It was also shown that assimilation of “accurate” synthetic interior data compensated for “inaccurate” synthetic boundary data. The time dependent calculations involved about 740 000 computational variables on a space–time grid. The annealing process was computed using a CM200 Connection Machine, and animated with a frame buffer. Local extrema in the penalty functional were seen to be associated with jagged “annealing ﬂaws” in the circulation ﬁelds. A careful annealing strategy led eventually to the global minimum and the correct, smooth circulation. The annealing samples in the ﬁnal stages yielded crude posterior error statistics. An HMC approach led directly to the global minimum but provided no error statistics. Those could be obtained, albeit crudely, by importancesampling near the minimum. HMC and other gradient methods become impractical for the estimation of smooth ﬁelds involving more than 106 gridded variables. In comparison, iterated representer methods have been successfully applied to the estimation of as many as 109 gridded variables modeling moderately nonlinear ﬂows, such as global weather, and climatic variability of the ocean–atmosphere: see Chapter 5. We do not, however, have efﬁcient data assimilation techniques for highly turbulent ﬂows marked by sharp fronts, outcrops or other neardiscontinuities.
Exercise 6.6.1 Construct a quartic polynomial J = J (u) having a graph like a “lopsided letter w”. Minimize J by simulated annealing.
6.7
Wellposedness in comoving domains
Oliger and Sundstr¨om (1978) established the uniqueness of solutions of the nonlinear shallowwater equations: ∂u + u · ∇u = −g∇h + F, ∂t ∂h + u · ∇h + h∇ · u = 0, ∂t
(6.7.1) (6.7.2)
184
6. Illposed forecasting problems
subject to the initial conditions h(x, 0) = h I (x) .
u(x, 0) = uI (x),
(6.7.3)
Let u(1) , h (1) and u(2) , h (2) be two solutions for the same forcing F and initial values uI , h I ; let u = u(1) − u(2) and h = h (1) − h (2) be the difference ﬁelds, and let 1 (1) 1 h u · u + g(h )2 (6.7.4) 2 2 be the total mechanical energy in the difference ﬁelds. Oliger and Sundstr¨om (1978) showed that ' ( ∂m + ∇ · u(1) m + gh (1) h u ∂t = gh u · ∇h (1) − gh u · ∇h (2) − g(h )2 ∇ · u(2) 1 + g(h )2 ∇ · u(1) − h (1) u · (u · ∇)u(2) . (6.7.5) 2 It follows that dE + A ≤ B E, (6.7.6) dt m=
where
E = E(t) = A = A(t) = B=
11 2
mdx,
(6.7.7)
ˆ + gh (1) h u · nˆ ds, u(1) · nm
(6.7.8)
D
B
1 max ∇ · u(1) , g/ h (i) 2 ∇h (i) i,x,t
(6.7.9)
and nˆ is the outward unit normal on the boundary B of the fixed domain D. Integrating (6.7.6) yields t E(t) ≤ E(0) −
e B(t−r ) A(r ) dr.
(6.7.10)
0
Note that the inverse timescale B is a constant. Now E(0) = 0 since the two solutions satisfy the same initial conditions. It has been shown in §3.3.4 that h (1) > 0, provided h I > 0. It is therefore clear that any boundary conditions ensuring A ≥ 0 also ensure E(t) ≡ 0, that is, the mixedinitialboundary value problem has a unique solution.
Exercise 6.7.1 It was shown in §6.2 that certain boundary conditions ensure uniqueness of solutions of the linear shallowwater equations. Verify that these same boundary conditions sufﬁce for the nonlinear equations.
6.7 Wellposedness in comoving domains
185
Exercise 6.7.2 (Bennett & Chua, 1999) Now suppose that the domain D consists of the same ﬂuid particles for all time t > 0, that is, the boundary B moves with the ﬂow, or is “comoving”. Prove that dE + C ≤ B E, dt where E and B are deﬁned as in (6.7.7) and (6.7.9), while C = C(t) ≡ gh (1) h u · nˆ ds.
(6.7.11)
(6.7.12)
B
Prove that uniqueness follows from any one boundary condition ensuring C ≥ 0, such as (i) h = 0 (ii) u · nˆ = 0 . (6.7.13) or 1 (iii) u · nˆ = k g/ h (1) 2 h The criticality of the ﬂow at the boundary is not an issue. Indeed, the local Froude (1) − 12 ˆ number u(1) · n(gh ) is effectively zero in the reference frame of the comoving boundary.
Exercise 6.7.3 How many conditions are needed in order to determine the motion of the boundary B?
References
Adams, R. A., 1975: Sobolev Spaces. Academic, New York, NY, 268 pp. Amodei, L., 1995: Solution approch´ee pour un probl`eme d’assimilation de donn´ees m´et´eorologiques avec prise en compte de l’erreur de mod`ele. Comptes Rendus de l’Acad´emie des Sciences, 321, S´erie IIa, 1087–1094.
Bennett, A. F., 1985: Array design by inverse methods. Progr. Oceanogr., 15, 129–156. Bennett, A. F., 1990: Inverse methods for assessing shipofopportunity networks and estimating circulation and winds from tropical expendable bathythermograph data. J. Geophys. Res., 95, 16 111–16 148.
Andersen, O. B., P. L. Woodworth, & R. A. Flather, 1995: Intercomparisons of recent ocean tide models. J. Geophys. Res., 100, 25 261–25 283.
Bennett, A. F., 1992: Inverse Methods in Physical Oceanography. Cambridge University Press, New York, NY (Reprinted 1999), 347 pp.
Anderson, D. L. T., J. Sheinbaum, & K. Haines, 1996: Data assimilation in ocean models. Rep. Progr. Phys., 59, 1209–1266.
Bennett, A. F., & J. R. Baugh, 1992: A parallel algorithm for variational assimilation in oceanography and meteorology. J. Atmos. Oc. Tech., 9, 426–433.
Azencott, R., 1992: Simulated Annealing. Wiley, New York, NY, 242 pp. Barth, N., & C. Wunsch, 1990: Oceanographic experiment design by simulated annealing. J. Phys. Oceanogr., 20, 1249–1263. Batchelor, G. K., 1973: An Introduction to Fluid Dynamics. Cambridge University Press, Cambridge, UK, 615 pp.
186
Bennett, A. F., & B. S. Chua, 1994: Openocean modeling as an inverse problem: the primitive equations. Mon. Wea. Rev., 122, 1326–1336. Bennett, A. F., & B. S. Chua, 1999: Open boundary conditions for Lagrangian geophysical ﬂuid dynamics. J. Comput. Phys., 153, 418–436.
References
Bennett, A. F., & P. E. Kloeden, 1981: The quasigeostrophic equations: approximation, predictability and equilibrium spectra of solutions. Q. J. R. Meteor. Soc., 107, 121–136. Bennett, A. F., & R. N. Miller, 1991: Weighting initial conditions in variational assimilation schemes. Mon. Wea. Rev., 119, 1004–1018. Bennett, A. F., & M. A. Thorburn, 1992: The generalized inverse of a nonlinear quasigeostrophic ocean circulation model. J. Phys. Oceanogr., 22, 213–230. Bennett, A. F., C. R. Hagelberg, & L. M. Leslie, 1992: Predicting hurricane tracks. Nature, 360, 423. Bennett, A. F., L. M. Leslie, C. R. Hagelberg, & P. E. Powers, 1993: Tropical cyclone prediction using a barotropic model initialized by a generalized inverse method. Mon. Wea. Rev., 121, 1714–1729. Bennett, A. F., B. S. Chua, & L. M. Leslie, 1996: Generalized inversion of a global numerical weather prediction model. Meteor. Atmos. Physics, 60, 165–178. Bennett, A. F., B. S. Chua, & L. M. Leslie, 1997: Generalized inversion of a global numerical weather prediction model, II: Analysis and implementation. Meteor. Atmos. Physics, 62, 129–140. Bennett, A. F., B. S. Chua, D. E. Harrison, & M. J. McPhaden, 1998: Generalized inversion of Tropical Atmosphere–Ocean (TAO) data and a coupled model of the tropical Paciﬁc. J. Climate, 11, 1768–1792. Bennett, A. F., B. S. Chua, D. E. Harrison, & M. J. McPhaden, 2000: Generalized inversion of Tropical Atmosphere–Ocean (TAO) data and a coupled model of the tropical Paciﬁc, II: The 1995–96 La Ni˜na and 1997–98 El Ni˜no. J. Climate, 13, 2770–2785.
187
Bleck, R., & L. T. Smith, 1990: A winddriven isopycnic coordinate model of the North and Equatorial Atlantic Ocean. 1. Model development and supporting experiments. J. Geophys. Res., 95, 3273–3285. Bogden, P., 2001: The impact of model–error correlations on regional data assimilative models and their observational arrays. J. Mar. Res., 59, 831–857. Bretherton, F. P., R. E. Davis, & C. B. Fandry, 1976: A technique for the objective analysis and design of oceanographic experiments applied to MODE73. DeepSea Research, 23, 559–582. Carton, J. A., G. Chepurin, & X. Cao, 2000a: A simple ocean data assimilation analysis of the global upper ocean 1950–95. Part I: Methodology. J. Phys. Oceanogr., 30, 294–309. Carton, J. A., G. Chepurin, & X. Cao, 2000b: A simple ocean data assimilation analysis of the global upper ocean 1950–95. Part II: Results. J. Phys. Oceanogr., 30, 311–326. Cartwright, D. E., 1999: Tides: A Scientiﬁc History. Cambridge University Press, Cambridge, UK, 272 pp. Chan, N. H., J. B. Kadane, R. N. Miller, & W. Palma, 1996: Estimation of tropical sea level anomaly by an improved Kalman ﬁlter. J. Phys. Oceanogr., 26, 1286–1303. Chandra, R., L. Dagum, D. Kohr, D. Maydan, J. McDonald, & R. Menon, 2001: Parallel Programming in OpenMP. Morgan Kaufmann, San Francisco, CA, 230 pp. Chapman, S., & T. G. Cowling, 1970: The Mathematical Theory of NonUniform Gases. Cambridge University Press, Cambridge, UK, 422 pp.
188
References
Chelton, D. B., J. C. Ries, B. J. Hanes, L.L. Fu, & P. S. Callahan, 2001: Satellite altimetry. In Satellite Altimetry and Earth Sciences, ed. L.L. Fu & A. Cazanave. Academic Press, San Diego, CA, 463 pp. Chua, B. S., & A. F. Bennett, 2001: An inverse ocean modeling system. Ocean Modelling, 3, 137–165. Cohn, S. E., 1997: An introduction to estimation theory. J. Met. Soc. Japan, 75, 257–288. Courant, R., & D. Hilbert, 1953: Methods of Mathematical Physics, vol. 1. Wiley Interscience, New York, NY, 560 pp. Courant, R., & D. Hilbert, 1962: Methods of Mathematical Physics, vol. 2. Wiley Interscience, New York, NY, 830 pp. Courtier, P., 1997: Dual formulation of fourdimensional assimilation. Q. J. R. Met. Soc., 123, 2449–2461. Courtier, P., J. Derber, R. Errico, J.F. Louis, & T. Vukicevic, 1993: Important literature on the use of adjoint, variational methods and the Kalman ﬁlter in meteorology. Tellus, 45A, 342–357.
Dee, D. P., & A. M. Da Silva, 1998: Data assimilation in the presence of forecast bias. Q. J. R. Meteor. Soc., 124, 269–295. Derber, J., & A. Rosati, 1989: A global oceanic data assimilation system. J. Phys. Oceanogr., 19, 1333–1347. Doodson, A. T., & H. D. Warburg, 1941: Admiralty Manual of Tides. Her Majesty’s Stationery Ofﬁce, London, UK, 270 pp. Dutz, J., 1998: Repression of fecundity in the neritic copepod Acartia clausi exposed to the toxic dinoﬂagellate Alexandrium lusitanicum: relationship between feeding and egg production. Mar. Ecol. Prog. Ser., 175, 97–107. Egbert, G. D., 1997: Tidal data inversion: interpolation and inference. Progr. Oceanogr., 40, 53–80. Egbert, G. D., & A. F. Bennett, 1996: Data assimilation methods for ocean tides, in Modern Approaches to Data Assimilation in Ocean Modeling, ed. P. MalanotteRizzoli. Elsevier, New York, NY, 147–179.
Cox, D. R., & D. V. Hinkley, 1974: Theoretical Statistics. Chapman and Hall, New York, NY, 511 pp.
Egbert, G. D., & R. D. Ray, 2000: Signiﬁcant dissipation of tidal energy in the deep ocean inferred from satellite altimeter data. Nature, 405, 775–778.
Cox, M. D., & K. Bryan, 1984: A numerical model of the ventilated thermocline. J. Phys. Oceanogr., 14, 674–687.
Egbert, G. D., & R. D. Ray, 2001: Estimates of M2 tidal energy dissipation from TOPEX/POSEIDON altimeter data. J. Geophys. Res., 106, 22 475–22 502.
Daley, R. A., 1991: Atmospheric Data Analysis. Cambridge University Press, Cambridge, UK, 457 pp.
Egbert, G. D., A. F. Bennett, & M. G. G. Foreman, 1994: TOPEX/POSEIDON tides estimated using a global inverse method. J. Geophys. Res., 99, 24 821– 24 852.
Daley, R., & E. Baker, 2000: The NAVDAS Source Book. NRL/PU/753000418. Naval Research Laboratory, Monterey, CA, 153 pp. Davidson, N. E., & B. J. McAvaney, 1981: The ANMRC tropical analysis scheme. Mon. Wea. Rev., 29, 155–168.
Eknes, M., & G. Evensen, 1997: Parameter estimation solving a weak constraint variational formulation for an Ekman model. J. Geophys. Res., 102, 12 479–12 491.
References
Elsberry, R. L., 1990: International experiments to study tropical cyclones in the western North Paciﬁc. Bull. Amer. Meteor. Soc., 71, 1305–1316. Emanuel, K., 1999: Thermodynamic control of hurricane intensity. Nature, 401, 665–669. Errico, R. M., 1997: What is an adjoint model? Bull. Amer. Met. Soc., 11, 2577–2750. Evensen, G., 1994: Sequential data assimilation with a nonlinear quasigeostrophic model using Monte Carlo methods to forecast error statistics. J. Geophys. Res., 99, 10 143–10 162. Ferziger, J. H., 1996: Large eddy simulation. In Simulation and Modeling of Turbulent Flows, ed. T. B. Gatski, M. Y. Hussaini, & J. L. Lumley. Oxford University Press, New York, NY, 314 pp. Fu, L. L., & I. Fukumori, 1996: A case study of the effects of errors in satellite altimetry on data assimilation, in Modern Approaches to Data Assimilation in Ocean Modeling, ed. P. MalanotteRizzoli. Elsevier, New York, NY, 77–96. Fukumori, I., 2001: Data assimilation for models. In Satellite Altimetry and Earth Sciences, ed. L.L. Fu & A. Cazanave. Academic Press, San Diego, CA, 463 pp. Fukumori, I., & P. MalanotteRizzoli, 1995: An approximate Kalman ﬁlter for ocean data assimilation; an example with an idealized Gulf Stream Model. J. Geophys. Res., 100, 6777–6793. Gelb, A. (ed.), 1974: Applied Optimal Estimation. MIT Press, Cambridge, MA, 374 pp.
189
of the upper tropical ocean. J. Comput. Phys., 81, 444–480. Ghil, M., K. Ide, A. Bennett, P. Courtier, M. Kimoto, M. Nagata, M. Saiki, & N. Sato (eds.), 1997: Data Assimilation in Meteorology and Oceanography, Theory and Practice. Meteorological Society of Japan, Tokyo, Japan, 496 pp. Giering, R. & T. Kaminsky, 1997: Recipes for adjoint code construction. ACM Trans. Math. Software, 24, 437–474. Gill, A. E., 1982: Atmosphere–Ocean Dynamics. Academic, London, UK, 662 pp. Golub, G. H., & C. F. Van Loan, 1989: Matrix Computations, 2nd ed. Johns Hopkins University Press, Baltimore, MD, 642 pp. Goerss, J. S., & P. A. Phoebus, 1993: The Multivariate Optimum Interpolation Analysis of Meteorological Data at the Fleet Numerical Oceanography Center. NRL/FR/7531929413. Naval Research Laboratory, Monterey, CA, 58 pp. Greiner, E., & C. Perigaud, 1994a: Assimilation of geosat altimetric data in a nonlinear reducedgravity model of the Indian Ocean. Part I: Adjoint approach and modeldata consistency. J. Phys. Oceanogr., 24, 1783–1803. Greiner, E., & C. Perigaud, 1994b: Assimilation of geosat altimetric data in a nonlinear reducedgravity model of the Indian Ocean. Part II: Some validation and interpretation of the assimilated results. J. Phys. Oceanogr., 26, 1735–1746.
Gelfand, I. M., & S. V. Fomin, 1963: Calculus of Variations. Prentice Hall, Englewood Cliffs, NJ, 232 pp.
Hackert, E. C., R. N. Miller, & A. J. Busalacchi, 1998: An optimized design for a moored instrument in the tropical Atlantic Ocean. J. Geophys. Res., 103, 7491–7509.
Gent, P. R., & M. A. Cane, 1989: A reduced gravity, primitive equation model
Hadamard, J., 1952: Lectures on Cauchy’s Problem in Linear Partial Differential
190
References
Equations. Dover Publications, New York, NY, 316 pp. Haltiner, G. J., & R. T. Williams, 1980: Numerical Prediction and Dynamical Meteorology, 2nd ed. John Wiley, New York, NY, 477 pp. Hoang, S., R. Baraille, O. Talagrand, X. Carton, & P. De Mey, 1997a: Adaptive ﬁltering: application to satellite data assimilation in oceanography. Dyn. Atmos. Oceans, 27, 257–281. Hoang, S., P. De Mey, O. Talagrand, & R. Baraille, 1997b: A new reducedorder adaptive ﬁlter for state estimation. Automatica, 33, 1475–1498. Hodur, R. M., 1997: The Naval Research Laboratory’s Coupled Ocean/Atmosphere Mesoscale Prediction System (COAMPS). Mon. Wea. Rev., 125, 1414–1430. Hogan, T. F., & T. E. Rosmond, 1991: The description of the U.S. Navy Operational Global Atmospheric Prediction System’s spectral forecast model. Mon. Wea. Rev., 119, 1786–1815. Holland, G. J., 1980: An analytical model of the wind and pressure proﬁles in hurricanes. Mon. Wea. Rev., 108, 1212–1218. Holton, J., 1992: An Introduction to Dynamical Meteorology, 3rd ed. Academic, San Diego, CA, 511 pp. Ide, K., P. Courtier, M. Ghil, & A. Lorenc, 1997: Uniﬁed notation for data assimilation: operational, sequential and variational. J. Met. Soc. Japan, 75, 181–189. Kasibhatla, P., M. Heimann, P. Rayner, N. Mahowald, R. G. Prinn, & D. E. Hartley (eds.), 2000: Inverse Methods in Global Biogeochemical Cycles. Geophysical Monograph 114, American Geophysical Union, Washington, D.C., 324 pp.
Keppenne, C. L., 2000: Data assimilation into a primitiveequation model with a parallel ensemble Kalman ﬁlter. Mon. Wea. Rev., 128, 1971–1981. Kessler, W. S., 1990: Observations of long Rossby waves in the northern tropical Paciﬁc. J. Geophys. Res., 95, 5183–5217. Kessler, W. S., & J. P. McCreary, Jr., 1993: The annual wind driven Rossby waves in the subthermocline equatorial Paciﬁc. J. Phys. Oceanogr., 23, 1192–1207. Kessler, W. S., M. C. Spillane, M. J. McPhaden, & D. E. Harrison, 1996: Scales of variability in the equatorial Paciﬁc inferred from the Tropical Atmosphere–Ocean buoy array. J. Climate, 9, 2999–3024. Kleeman, R. C., A. M. Moore, & N. R. Smith, 1995: Assimilation of subsurface thermal data into a simple ocean model for the initialization of an intermediate tropical coupled ocean–atmosphere forecast model. Mon. Wea. Rev., 123, 3101–3113. Klinker, E., F. Rabier, G. Kelly, & J.F. Mahfouf, 2000: The ECMWF operational implementation of fourdimensional variational assimilation. III: Experimental results and diagnostics with operational conﬁguration. Q. J. R. Meteorol. Soc., 126, 1191–1215. Kruger, J., 1993: Simulated annealing: A tool for data assimilation into an almost steady model state. J. Phys. Oceanogr., 23, 679–688. Kundu, P. K., 1990: Fluid Mechanics. Academic, San Diego, CA, 638 pp. Ladyzhenskaya, O. A., 1969: The Mathematical Theory of Viscous Incompressible Flow, 2nd English ed., translated from the Russian by R. A. Silverman & J. Chu. Gordon & Breach, New York, NY, 224 pp.
References
Lagarde, T., A. Piacentini, & O. Thual, 2001: A new representation of dataassimilation methods: The PALM ﬂowcharting approach. Q. J. R. Meteorol. Soc., 127, 189–207. Lanczos, C., 1966: The Variational Principles of Mechanics, 3rd ed. University of Toronto, Toronto, Canada, 375 pp. Le Dimet, F., & O. Talagrand, 1986: Variational algorithms for analysis and assimilation of meteorological observations. Tellus, 38A, 97–110. Le Dimet, F.X., H.E. Ngodock, B. Luong, & J. Verron, 1997: Sensitivity analysis in variational data assimilation. J. Met. Soc. Japan, 75, 245–255. Le Provost, C., 2001: Ocean tides. In Satellite Altimetry and Earth Sciences, ed. L.L. Fu & A. Cazanave. Academic Press, San Diego, CA, 463 pp. Le Provost, C. L., & F. Lyard, 1997: Energetics of the barotropic ocean tides: an estimate of bottom friction dissipation from a hydrodynamic model. Prog. Oceanogr., 40, 37–52. Le Provost, C. L., M. L. Genco, F. Lyard, P. Vincent, & P. Canceil, 1994: Spectroscopy of the world ocean tides from a hydrodynamic model. J. Geophys. Res., 99, 24 777–24 797. Le Provost, C., A. F. Bennett, & D. E. Cartwright, 1995: Ocean tides for and from TOPEX/POSEIDON. Science, 267, 639–642. Lewis, J. M., & J. C. Derber, 1985: The use of adjoint equations to solve a variational adjustment problem with advective constraints. Tellus, 37A, 309–322. Lions, J. L., 1971: Optimal Control of Systems Governed by Partial Differential Equations. SpringerVerlag, Berlin, Federal Republic of Germany, 396 pp.
191
Ljung, L., & T. S¨oderstr¨om, 1987: Theory and Practice of Recursive Identiﬁcation. MIT Press, Cambridge, MA, 529 pp. Lorenc, A., 1997: Atmospheric data assimilation. Forecasting Research Scientiﬁc Paper 34, Meteorological Ofﬁce, Bracknell, UK. Lorenc, A. C., S. P. Ballard, R. S. Bell, N. B. Ingleby, P. L. F. Andrews, D. M. Barker, J. R. Bray, A. M. Clayton, T. Dalby, D. Li, T. J. Payne, & F. W. Saunders, 2000: The Met. Ofﬁce global threedimensional variational data assimilation scheme. Q. J. R. Meteorol. Soc., 126, 2991–3012. Lyard, F. H., 1999: Data assimilation in a wave equation: a variational approach for the Grenoble tidal model. J. Comp. Phys., 149, 1–31. Lyard, F. H., & C. Le Provost, 1997: Energy budget of the tidal hydrodynamic model FES 94.1. Geophys. Res. Lett., 24, 687–690. Mahfouf, J.F., & F. Rabier, 2000: The ECMWF operational implementation of fourdimensional variational assimilation. II: Experimental results with improved physics. Q. J. R. Meteorol. Soc., 126, 1171–1190. MalanotteRizzoli, P. (ed.), 1996: Modern Approaches to Data Assimilation in Ocean Modeling. Elsevier, Amsterdam, The Netherlands, 445 pp. MalanotteRizzoli, P., I. Fukumori, & R. E. Young, 1996: A methodology for the construction of a hierarchy of Kalman ﬁlters for nonlinear primitive equation models. In Modern Approaches to Data Assimilation in Ocean Modeling, ed. P. MalanotteRizzoli. Elsevier, New York, NY, 297–317. Marotske, J., R. Giering, K. Q. Zhang, D. Stammer, C. Hill, & T. Lee, 1999: Construction of the adjoint MIT ocean
192
References
general circulation model and application to Atlantic heat transport sensitivity. J. Geophys. Res., 104, 29 529–29 547. Marshall, J., C. Hill, L. Perelman, & A. Adcroft, 1997a: Hydrostatic, quasihydrostatic, and nonhydrostatic ocean modeling. J. Geophys. Res., 102, 5733–5752. Marshall, J., A. Adcroft, C. Hill, L. Perelman, & C. Heisey, 1997b: A ﬁnitevolume, incompressible Navier–Stokes model for studies of the ocean on parallel computers. J. Geophys. Res., 102, 5753–5766. McIntosh, P. C., & A. F. Bennett, 1984: Open ocean modeling as an inverse problem: M2 tides in Bass Strait. J. Phys. Oceanogr., 14, 601–614. McPhaden, M. J., 1993: TOGA–TAO and the 1991–93 El Ni˜no Southern Oscillation event. Oceanography, 6, 36–44. McPhaden, M. J., 1999a: Genesis and evolution of the 1997–98 El Ni˜no. Science, 283, 950–954.
Miller, R. N., E. D. Zaron, & A. F. Bennett, 1994: Data assimilation in models with convective adjustment. Mon. Wea. Rev., 122, 2607–2613. Miller, R. N., A. J. Busalacchi, & E. C. Hackert, 1995: Sea surface topography ﬁelds of the tropical Paciﬁc from data assimilation. J. Geophys. Res., 100, 13 389–13 425. Muccino, J. C., & A. F. Bennett, 2001: Generalized inversion of the Korteweg–deVries equation. Dyn. Atmos. Oceans, (to appear). Oliger, J., & A. Sundstr¨om, 1978: Theoretical and practical aspects of some initial boundary value problems in ﬂuid dynamics, SIAM J. App. Math., 35, 419–446. Pacanowski, R. C., & S. G. H. Philander, 1981: Parameterization of vertical mixing in numerical models of tropical oceans. J. Phys. Oceanogr., 11, 1443–1451.
McPhaden, M. J., 1999b: El Ni˜no: the child prodigy of 1997–98. Nature, 398, 559–562.
Pacheco, P., 1996: Parallel Programming with MPI. Morgan Kaufmann, San Francisco, CA, 440 pp.
Meditch, J. S., 1970: On state estimation for distributed parameter systems. J. Franklin Inst., 290, 45–59.
Parker, R. L., 1994: Geophysical Inverse Theory. Princeton University Press, Princeton, NJ, 386 pp.
Mesinger, F., & A. Arakawa, 1976: Numerical Methods Used in Atmospheric Models. GARP Publication Series No. 14, WMO/ICSU Joint Organizing Committee, Geneva, Switzerland, 64 pp.
Passi, R. M., R. K. Goodrich, & J. C. Derber, 1996: A convolution algorithm with application to data assimilation. SIAM J. Sci. Comput., 17, 942–955.
Metropolis, N., A. W. Rosenbluth, M. N. Rosenbluth, A. H. Teller, & E. Teller, 1953: Equation of state calculation by fast computing machines. J. Chem. Physics, 21, 1087–1092.
Pedlosky, J., 1987: Geophysical Fluid Dynamics, 2nd ed. SpringerVerlag, New York, NY, 710 pp.
Miller, R. N., 1996: Introduction to the Kalman Filter. ECMWF Workshop on Data Assimilation, Reading, UK.
Press, W. H., S. A. Teukolsky, W. T. Vetterling, & B. P. Flannery, 1986: Numerical Recipes, the Art of Scientiﬁc Computing. Cambridge University Press, New York, NY, 735 pp.
References
Rabier, F., H. J¨arvinen, E. Klinker, J.F. Mahfouf, & A. Simmons, 2000: The ECMWF operational implementation of fourdimensional assimilation. I: Experimental results with simpliﬁed physics. Q. J. R. Meteorol. Soc., 126, 1143–1170. Reichle, R. H., D. Entekhabi, & D. B. McLaughlin, 2001: Downscaling of radiobrightness measurements for soil moisture estimation: a fourdimensional variational data assimilation approach. Water Resource Res., 37, 2353–2364. Reichle, R. H., D. B. McLaughlin, & D. Entekhabi, 2002: Hydrologic data assimilation with the Ensemble Kalman ﬁlter. Mon. Wea. Rev., 130, 103–114. Reid, W. T., 1968: Generalized inverses of differential and integral operators. In Theory and Applications of Generalized Inverses of Matrices, ed. T. L. Boullion & P. L. Odell, pp. 1–25. Texas Technical College, Lubbock, TX. Reynolds, R. W., & T. M. Smith, 1995: A high resolution global sea surface temperature climatology. J. Climate, 8, 1571–1583. Rodgers, C. D., 2000: Inverse Methods for Atmospheric Sounding: Theory and Practice. World Scientiﬁc, Singapore, 256 pp. Sasaki, Y., 1970: Some basic formalisms in numerical variational analysis. Mon. Wea. Rev., 98, 875–883. Schiff, L. I., 1949: Quantum Mechanics. McGrawHill, New York, NY, 404 pp. Schrama, E. J. O., & R. Ray, 1994: A preliminary tidal analysis of TOPEX/POSEIDON altimetry. J. Geophys. Res., 99, 24 799–24 808. Shum, C. K., P. L. Woodworth, O. B. Andersen, G. Egbert, O. Francis, C.
193
King, S. Klosko, C. Le Provost, X. Li, J. M. Molines, M. Parke, & C. Wunsch, 1997: Accuracy assessment of recent ocean tidal models. J. Geophys. Res., 102, 25 173–25 194. Soreide, N. N., D. C. McClurg, W. H. Zhu, M. J. McPhaden, D. W. Denbo, & M. W. Renton, 1996: World Wide Web access to realtime and historical data from the TAO array of moored buoys in the tropical Paciﬁc Ocean: updates for 1996. Proc. OCEANS 96, Fort Lauderdale, FL, MTS/IEEE, 1354–1359. Stakgold, I., 1979: Green’s Functions and Boundary Value Problems. Wiley, New York, NY, 638 pp. Stammer, D., R. Davis, L.L. Fu, I. Fukumori, R. Giering, T. Lee, J. Marotzke, J. Marshall, D. Menemenlis, P. Niiler, C. Wunsch, & V. Zlotnicki, 2000: Ocean state estimation in support of CLIVAR and GODAE. CLIVAR Exchanges, 5, 3–5. www.clivar.org Thacker, W. C., & R. B. Long, 1988: Fitting dynamics to data. J. Geophys. Res., 93, 1227–1240. Thi´ebaux, H. J., & M. A. Pedder, 1987: Spatial Objective Analysis. Academic, New York, NY, 299 pp. Uboldi, F., & M. Kamachi, 2000: Time–space weakconstraint data assimilation for nonlinear models. Tellus, 52A, 412–421. Velden, C. S., C. M. Hayden, W. P. Menzel, J. L. Franklin, & J. S. Lynch, 1992: The impact of satellitederived winds on numerical hurricane track forecasts. Wea. & Forecasting, 7, 107–118. Wahba, G., & J. Wendelberger, 1980: Some new mathematical methods for variational objective analysis using splines and crossvalidation. Mon. Wea. Rev., 108, 1122–1143.
194
References
Wallace, J. M., & P. V. Hobbs, 1977: Atmospheric Science: An Introductory Survey. Academic, New York, NY, 467 pp. Weinert, H. L. (ed.), 1982: Reproducing Kernel Hilbert Spaces, Applications in Statistical Signal Processing. Hutchinson Ross, Stroudsburg, PA, 654 pp. Woodruff, S. D., R. J. Slutz, R. L. Jenne, & P. Steurer, 1987: A comprehensive atmosphere–ocean data set. Bull. Amer. Meteor. Soc., 68, 1239–1250.
algorithm to a simple transport problem. Tellus, 52A, 109–128. Xu, Q., & J. Gao, 1999: Generalized adjoint for physical processes with parameterized discontinuities. Part IV: Minimization problems in multidimensional space. J. Atmos. Sci., 56, 994–1002. Yoshida, K., 1980: Functional Analysis, 6th ed. SpringerVerlag, Berlin, Federal Republic of Germany, 496 pp.
Wunsch, C., 1996: The Ocean Circulation Inverse Problem. Cambridge University Press, New York, NY, 442 pp.
Zebiak, S. E., & M. A. Cane, 1987: A model El Ni˜no/Southern Oscillation. Mon. Wea. Rev., 115, 2262–2278.
Wunsch, C., 1998: The work done by the wind on the ocean circulation. J. Phys. Oceanogr., 28, 2331–2339.
Zhang, S., X. Zou, J. Ahlquist, & I. M. Navon, 2000: Use of differentiable and nondifferentiable optimization algorithms for variational data assimilation with discontinuous cost functions. Mon. Wea. Rev., 128, 4031–4044.
Xu, L., & R. Daley, 2000: Towards a true 4dimensional data assimilation algorithm: application of a cycling representer
Appendix A Computing exercises
These computational exercises complement the analytical development of variational data assimilation in the text, and also serve to develop the conﬁdence needed for more ambitious calculations. All code for these exercises is available at an anonymous ftp site. The linear, onedimensional “toy” model of §1.1 is upgraded here to a linear, twodimensional shallowwater model in both continuous and ﬁnitedifference form. Continuous and discrete penalty functionals are developed, and the respective Euler–Langrange equations are derived. Representers are calculated directly, so the generalized inverse may then be calculated directly or indirectly.
A.1
Forward model
A.1.1
Preamble
The exercise is to construct and run a simple forward model using standard numerical methods. You may obtain the source code from our anonymous ftp site (ftp ftp.oce.orst.edu, then cd /dist/bennett/class), along with some plotting utilities. It is assumed that your computer environment is UNIX, with the Fortran compilation command ‘f 77’.
195
196
Appendix A Computing exercises
Figure A.1.1 Periodic channel with rigid walls, rotating at the rate 2f about an axis normal to the x yplane.
Y v
0 u,v,q periodic in x
u,v,q periodic in x
y
v
0
0 X
0
A.1.2
x
Model
We consider a linear shallowwater model ∂u ∂q − fv+g + ru u = Fu , ∂t ∂x ∂v ∂q + fu +g + rv v = Fv , ∂t ∂y ∂q ∂u ∂v +H + + rq q = 0, ∂t ∂x ∂y on the domain 0 ≤ x ≤ X and 0 ≤ y ≤ Y (see Fig. A.1.1). A.1.3
Initial conditions
The initial values are u(x, y, 0) = I u (x, y) = 0, v(x, y, 0) = I v (x, y) = 0, and q(x, y, 0) = I q (x, y) = 0. A.1.4
Boundary conditions
The north and south walls are rigid: v(x, 0, t) = v(x, Y, t) = 0, while all ﬁelds are periodic in the xdirection: u(x ± X, y, t) = u(x, y, t), v(x ± X, y, t) = v(x, y, t), and q(x ± X, y, t) = q(x, y, t).
A.1 Forward model
A.1.5
197
Model forcings
The model forcings are Fu = −Cd ρa u a 2 /(H ρw ), and Fv = 0. A.1.6
Model parameters
The following parameters are suggested: zonal period, X = 2000 km meridional width, Y = 1000 km mean depth, H = 5000 m time interval, T = 1.8 × 104 s gravitational acceleration, g = 9.806 m s−2 Coriolis parameter, f = 1.0 × 10−4 s−1 damping coefﬁcient, ru = (1.8 × 104 s)−1 damping coefﬁcient, rv = (1.8 × 104 s)−1 damping coefﬁcient, rq = (1.8 × 104 s)−1 drag coefﬁcient, Cd = 1.6 × 10−3 air density, ρa = 1.275 kg m−3 water density, ρw = 1.0 × 103 kg m−3 zonal wind, u a = 5 m s−1 .
A.1.7
Numerical model
The differential equations are discretized on the Arakawa Cgrid (see Fig. A.1.2) with a forward–backward scheme for timestepping (Mesinger & Arakawa, 1976) given as follows: k k k u i+1, vi,k j+1 − vi,k j qi,k+1 j − qi, j j − u i, j +H + + rq qi,k j = 0, t x y k k k vi,k j+1 + vi,k j + vi−1, u i,k+1 j − u i, j j+1 + vi−1, j − f t 4 k+1 k+1 qi, j − qi−1, j +g + ru u i,k j = Fu i,k j , x k k k k k u i+1, vi,k+1 j − vi, j j + u i, j + u i+1, j−1 + u i, j−1 + f t 4 k+1 qi, j − qi,k+1 j−1 +g + rv vi,k j = Fvi,k j , y
198
Appendix A Computing exercises
Spatial: qi,j+1 vi,j+1
vi1,j+1
j ui,j
∆y
qi,j
ui+1,j
qi1,j
i
qi+1,j
vi1,j
vi,j
ui,j1
qi,j1
ui+1,j1
∆x
Temporal: U0
U1
U k1
Uk
U NK1
U NK k
λ
λ
1
λ
k
NK
∆t λu
u U= v q
v λ = λq
λ
Figure A.1.2 Arakawa Cgrid for space differences; staggering of forward and adjoint variables for time differences.
where qi,k j : i = 1, NI ;
j = 1, NJ − 1;
k = 0, NK − 1,
: i = 1, NI ;
j = 1, NJ − 1;
k = 0, NK − 1,
vi,k j : i = 1, NI ;
j = 2, NJ − 1;
k = 0, NK − 1.
u i,k j and
Rigid boundary conditions k vi,1 = 0,
and k vi,N J = 0.
Periodic boundary conditions u k0, j = u kN I, j , u kN I +1, j = u k1, j , k k v0, j = v N I, j , k v kN I +1, j = v1, j, k k q0, j = q N I, j ,
A.2 Variational data assimilation
199
and k q Nk I +1, j = q1, j.
A.1.8
Numerical model parameters
The following parameters are suggested: number of grid points in xdirection, N I = 20 number of grid points in ydirection, N J = 11 number of time steps, N K = 100 grid spacing (xdirection), x = 100 km grid spacing (ydirection), y = 100 km time step, t = 180 s.
A.1.9
Numerical code and output
Source code: fwd.f To compile the source code: f77O3 fwd.f o fwd Output ﬁle: ufwd.dat, vfwd.dat and qfwd.dat
A.1.10
To generate a plot
(i) Postprocessing Source code: postprocess.f To compile the source code: f 77 postprocess.f o postprocess Input ﬁle: ufwd.dat, vfwd.dat and qfwd.dat Output ﬁle: u.dat, v.dat, q.dat, uy.dat, vy.dat, qy.dat (ii) Line plot gnuplot line.gnu (input ﬁle: qy.dat; output ﬁle: qy.ps) (iii) Contour plot gnuplot contour.gnu (input ﬁle: q.dat; output ﬁle: q.ps)
A.2
Variational data assimilation
A.2.1
Preamble
The exercise is to reformulate the forward model of Exercise A.1 as an inverse model, to deﬁne a penalty functional, to derive the associated system of Euler–Lagrange (EL) equations and to express its solution using representers.
200
A.2.2
Appendix A Computing exercises
Penalty functional
Formulate the penalty functional J for the following problem (see §1.2, Inverse models):
Equations of motion ∂u ∂q − fv+g + ru u = Fu + f u , ∂t ∂x ∂v ∂q + fu +g + rv v = Fv + f v , ∂t ∂y ∂q ∂v ∂u +H + + rq q = f q . ∂t ∂x ∂y
Initial conditions u(x, y, 0) = I u (x, y) + i u (x, y), v(x, y, 0) = I v (x, y) + i v (x, y), q(x, y, 0) = I q (x, y) + i q (x, y).
Rigid boundary conditions v(x, 0, t) = b0 (x, t), v(x, Y, t) = bY (x, t).
Periodic boundary conditions u(x ± X, y, t) = u(x, y, t), v(x ± X, y, t) = v(x, y, t), q(x ± X, y, t) = q(x, y, t).
Data dm = q(xm , ym , tm ) + m , A.2.3
1 ≤ m ≤ M.
Euler–Lagrange equations
Derive EL equations for the extremum of the penalty functional J . A.2.4
Representer solution
Solve the EL equations using representers (see §1.3: Solving the Euler–Lagrange equations using representers).
A.2 Variational data assimilation
201
Solutions for A.2.2–A.2.4
A.2.5
Penalty functional, in terms of residuals: J = J [u, v, q] T =
W uf
X
Y
dt 0
dx 0
q
+ Wbv
T
W vf
dt
Y
dt 0
0
Y
0
0
T
X
dy(i v (x, y))2 + Wi
d x(b0 (x, t)) + 2
dt 0
0
X
Wbv
0
0
dy(i u (x, y))2
dx 0
Y dy(i q (x, y))2
dx 0
0
T
X d x(bY (x, t))2 + w
dt 0
dy( f v (x, y, t))2
Y
0
q
dx
Y dx
X dy( f q (x, y, t))2 + Wiu
dx
X
X
0
X
0
+ Wiv
2
0
T +Wf
dy( f (x, y, t)) + u
M m=1
0
Penalty functional, in terms of state variables: J = J [u, v, q] T =
W uf
X dt
0
+ W vf
0
dy 0
X dt
dx
dy
0
0
T
X
Y
dt
dx
0
0
X
Y
+ Wiu 0
0
X
Y dx
∂u ∂v + ∂x ∂y
dy{v(x, y, 0) − I v (x, y)}2
Y dy {q(x, y, 0) − I q (x, y)}2
dx 0
dy 0
∂q +H ∂t
0
X q
+ Wi
0
2
∂q ∂v + fu +g + rv v − Fv ∂t ∂y
dy{u(x, y, 0) − I u (x, y)}2
dx
0
∂u ∂q − fv+g + ru u − Fu ∂t ∂x
Y
0
q +Wf
+ Wiv
dx
T
Y
2
2 + rq q
(m )2 .
202
Appendix A Computing exercises
+ Wbv
+ Wbv
T
X d x{v(x, 0, t)}2
dt 0
0
T
X d x{v(x, Y, t)}2
dt 0
+w
M
0
{q(xm , ym , tm ) − dm }2 .
m=1
Weighted residuals
∂ uˆ ∂ qˆ λ ≡ − f vˆ + g + ru uˆ − Fu , ∂t ∂x ∂ vˆ ∂ qˆ λv ≡ W vf + f uˆ + g + rv vˆ − Fv , ∂t ∂y ∂ qˆ ∂ uˆ ∂ vˆ q q λ ≡ Wf +H + + rq qˆ . ∂t ∂x ∂y u
W uf
Euler–Lagrange equations ∂λu ∂λq + f λv − H + ru λu = 0, ∂t ∂x ∂λv ∂λq − − f λu − H + rv λv = 0, ∂t ∂y u ∂λq ∂λ ∂λv − −g + + rq λq ∂t ∂x ∂y −
= −w
M
ˆ m , ym , tm ) − dm ) δ(x − xm )δ(y − ym )δ(t − tm ). (q(x
m=1
λu (x, y, T ) = 0, λv (x, y, T ) = 0, λq (x, y, T ) = 0. λv (x, 0, t) = 0, λv (x, Y, t) = 0. λu (x ± X, y, t) = λu (x, y, t), λv (x ± X, y, t) = λv (x, y, t), λq (x ± X, y, t) = λq (x, y, t).
A.2 Variational data assimilation
' (−1 u ∂ uˆ ∂ qˆ λ , − f vˆ + g + ru uˆ = Fu + W uf ∂t ∂x ' (−1 v ∂ vˆ ∂ qˆ λ , + f uˆ + g + rv vˆ = Fv + W vf ∂t ∂y ' q (−1 q ∂ qˆ ∂ uˆ ∂ vˆ λ . +H + + rq qˆ = W f ∂t ∂x ∂y ' (−1 u ˆ u(x, y, 0) = I u (x, y) + Wiu λ (x, y, 0), ' v (−1 v v λ (x, y, 0), vˆ (x, y, 0) = I (x, y) + Wi ' ( −1 q ˆ q(x, y, 0) = I q (x, y) + Wi λq (x, y, 0). ' (−1 q vˆ (x, 0, t) = H Wbv λ (x, 0, t), ' v (−1 q vˆ (x, Y, t) = −H Wb λ (x, Y, t). ˆ ± X, y, t) = u(x, ˆ u(x y, t), vˆ (x ± X, y, t) = vˆ (x, y, t), ˆ ± X, y, t) = q(x, ˆ q(x y, t).
Firstguess ∂u F ∂q F − f vF + g + ru u F = Fu , ∂t ∂x ∂v F ∂q F + f uF + g + rv v F = Fv , ∂t ∂y ∂q F ∂u F ∂v F +H + + rq q F = 0. ∂t ∂x ∂y u F (x, y, 0) = I u (x, y), v F (x, y, 0) = I v (x, y), q F (x, y, 0) = I q (x, y). v F (x, 0, t) = 0, v F (x, Y, t) = 0. u F (x ± X, y, t) = u F (x, y, t), v F (x ± X, y, t) = v F (x, y, t), q F (x ± X, y, t) = q F (x, y, t).
203
204
Appendix A Computing exercises
Representer adjoint equations ∂αmu ∂α q + f αmv − H m + ru αmu = 0, ∂t ∂x q ∂αmv ∂α − − f αmu − H m + rv αmv = 0, ∂t ∂y u q ∂α ∂αm ∂α v − m −g + m + rq αmq = δ(x − xm )δ(y − ym )δ(t − tm ). ∂t ∂x ∂y
−
αmu (x, y, T ) = 0, αmv (x, y, T ) = 0, αmq (x, y, T ) = 0. αmv (x, 0, t) = 0, αmv (x, Y, t) = 0. αmu (x ± X, y, t) = αmu (x, y, t), αmv (x ± X, y, t) = αmv (x, y, t), αmq (x ± X, y, t) = αmq (x, y, t).
Representer equations ' (−1 u ∂rmu ∂r q αm , − f rmv + g m + ru rmu = W uf ∂t ∂x ' (−1 v ∂rmv ∂r q αm , + f rmu + g m + rv rmv = W vf ∂t ∂y u ' q (−1 q ∂rmq ∂rm ∂rmv αm . + rq rmq = W f +H + ∂t ∂x ∂y ' (−1 u rmu (x, y, 0) = Wiu αm (x, y, 0), ' ( −1 v rmv (x, y, 0) = Wiv αm (x, y, 0), ' ( q −1 q rmq (x, y, 0) = Wi αm (x, y, 0). ' (−1 q rmv (x, 0, t) = H Wbv αm (x, 0, t), ' ( −1 q rmv (x, Y, t) = −H Wbv αm (x, Y, t). rmu (x ± X, y, t) = rmu (x, y, t), rmv (x ± X, y, t) = rmv (x, y, t), rmq (x ± X, y, t) = rmq (x, y, t).
A.3 Discrete formulation
205
Extremum of J M
ˆ u(x, y, t) = u F (x, y, t) +
βˆ m rmu (x, y, t),
m=1
vˆ (x, y, t) = v F (x, y, t) +
M
βˆ m rmv (x, y, t),
m=1
ˆ q(x, y, t) = q F (x, y, t) +
M
βˆ m rmq (x, y, t),
m=1
M q rlm + w−1 δlm βˆ l = h m
(m = 1, M),
l=1
where rlm = rl (xm , ym , tm ), and h m = dm − q F (xm , ym , tm ).
A.3
Discrete formulation
A.3.1
Preamble
Verify the discrete formulation given here in detail. Derive the corresponding equations for the representers and their adjoints. Compare with the source code rep.f. A.3.2
Penalty functional J = J [u, v, q] =
W uf
N K −1 N J −1 NI k=0
+ W vf
j=1
q
+ Wiv
i=1
j=2
N K −1 N J −1 k=0
+ Wiu
2 xyt ( f u )i,k+1 j
N K −1 N J −1 NI k=0
+Wf
j=1
N J −1
NI
j=1
i=1
NI NJ j=1 i=1
2 ( f v )i,k+1 xyt j
i=1 NI
2 xyt ( f q )i,k+1 j
i=1
((i u )i, j )2 xy
((i v )i, j )2 xy
Appendix A Computing exercises
206
q
+ Wi
+ Wbv
N J −1
NI
j=1
i=1
((i q )i, j )2 xy
NK NI
2 (b0 )ik xt
k=1 i=1
+ Wbv
NK NI
2 (bY )ik xt
k=1 i=1 M
+w
(m )2 ,
m=1
where (
f u )i,k+1 j
=
+g
(
f v )i,k+1 j
=
(
=
k k vi,k j+1 + vi,k j + vi−1, j+1 + vi−1, j
− f t k+1 k+1 qi, j − qi−1, j x
k vi,k+1 j − vi, j
+g f q )i,k+1 j
k u i,k+1 j − u i, j
k qi,k+1 j − qi, j
t
+H
4 + ru u i,k j − (F u )i,k j ,
k k k k u i+1, j + u i, j + u i+1, j−1 + u i, j−1
+ f t k+1 qi, j − qi,k+1 j−1 y
4 + rv vi,k j − (F v )i,k j ,
k k u i+1, j − u i, j
x
+
vi,k j+1 − vi,k j y
+ rq qi,k j ,
(i u )i, j = u i,0 j − (I u )i, j , (i v )i, j = vi,0 j − (I v )i, j , (i q )i, j = qi,0 j − (I q )i, j , k (b0 )ik = vi,1 , k (bY )ik = vi,N J,
m = qikmm, jm − dm . A.3.3
Weighted residuals
(λu )i,k+1 j
≡
W uf +g
ˆ i,k j uˆ i,k+1 j −u t
− f
k+1 ˆ i−1, qˆ i,k+1 j −q j
x
k k ˆ i−1, vˆ i,k j+1 + vˆ i,k j + vˆ i−1, j+1 + v j
4
+ ru uˆ i,k j − (F u )i,k j
,
A.3 Discrete formulation
(λv )i,k+1 j
≡
W vf +g
(λq )i,k+1 j
≡
q Wf
vˆ i,k+1 ˆ i,k j j −v t
+ f
ˆ i,k+1 qˆ i,k+1 j −q j−1 y ˆ i,k j qˆ i,k+1 j −q t
k k ˆ i,k j + uˆ i+1, ˆ i,k j−1 uˆ i+1, j +u j−1 + u
4
+ rv vˆ i,k j − (F v )i,k j
+H
k ˆ i,k j uˆ i+1, j −u
x
+
, vˆ i,k j+1 − vˆ i,k j
y
+ rq qˆ i,k j
Euler–Lagrange equations v k+1 v k+1 v k+1 u k (λv )i,k+1 (λu )i,k+1 j − (λ )i, j j+1 + (λ )i, j + (λ )i−1, j+1 + (λ )i−1, j − + f t 4 q k+1 (λq )i,k+1 j − (λ )i−1, j −H + ru (λu )i,k+1 j = 0, x k+1 u k+1 u k+1 u k+1 v k (λu )i+1, (λv )i,k+1 j + (λ )i, j + (λ )i+1, j−1 + (λ )i, j−1 j − (λ )i, j − − f t 4 q k+1 (λq )i,k+1 j − (λ )i, j−1 −H + rv (λv )i,k+1 j = 0, y k u k q k (λu )i+1, (λv )i,k j+1 − (λv )i,k j (λq )i,k+1 j − (λ )i, j j − (λ )i, j − −g + t x y + rq (λq )i,k+1 j = −w
M δi,im δ j, jm δk,km k qˆ im , jm − dm , xyt m=1
where k = 1, N K − 1. (λu )i,N jK t (λv )i,N jK t (λq )i,N jK t
= 0, = 0, −g
NK u NK (λu )i+1, j − (λ )i, j
x
= −w
+
K − (λv )i,N jK (λv )i,N j+1
y
M δi,im δ j, jm δ N K ,km N K qˆ im , jm − dm . xyt m=1
k (λv )i,1 = 0, k (λv )i,N J = 0
207
(computational boundary conditions).
.
Appendix A Computing exercises
208
(λu )k0, j = (λu )kN I, j , (λu )kN I +1, j = (λu )k1, j , (λv )k0, j = (λv )kN I, j , (λv )kN I +1, j = (λv )k1, j , (λq )k0, j = (λq )kN I, j , (λq )kN I +1, j = (λq )k1, j . ˆ i,k j uˆ i,k+1 j −u t
vˆ i,k+1 ˆ i,k j j −v t
ˆ i,k j qˆ i,k+1 j −q t Wiu
−f
k k ˆ i−1, vˆ i,k j+1 + vˆ i,k j + vˆ i−1, j+1 + v j
4 ' (−1 u k+1 + W uf (λ )i, j ,
uˆ i,0 j − (I u )i, j
=
vˆ i,0 j − (I v )i, j
=
t
t
qˆ i,0 j − (I q )i, j
t '
( v −1
'
( v −1
k vˆ i,1 = Wb
k vˆ i,N J = Wb
=
− f
+g
k+1 ˆ i−1, qˆ i,k+1 j −q j
x
t
+ f
t f
y
f
4
,
x
1 u 1 u 1 u 1 (λu )i+1, j + (λ )i, j + (λ )i+1, j−1 + (λ ) i, j−1
(λq )i,1 j − (λq )i,1 j−1
(λq )i,1 j
y
1 v 1 (λv )i,1 j+1 + (λv )i,1 j + (λv )i−1, j+1 + (λ )i−1, j
1 (λq )i,1 j − (λq )i−1, j
(λv )i,1 j
+H q Wi
(λu )i,1 j
+H
+ ru uˆ i,k j = (Fu )i,k j k+1 k k ˆ i,k j + uˆ i+1, ˆ i,k j−1 uˆ i+1, qˆ i, j − qˆ i,k+1 j +u j−1 + u j−1 +f +g 4 y ' ( −1 + rv vˆ i,k j = (Fv )i,k j + W vf (λv )i,k+1 j , k ' q (−1 q k+1 ˆ i,k j uˆ i+1, vˆ i,k j+1 − vˆ i,k j j −u (λ )i, j . +H + + rq qˆ i,k j = W f x y
t
Wiv
4
− rv (λv )i,1 j ,
y − rq (λq )i,1 j . k k + (λu )i+1,1 (λu )i,1 4
+H
k u k (λu )i,N J −1 + (λ )i+1,N J −1 4
uˆ k0, j = uˆ kN I, j , uˆ kN I +1, j = uˆ k1, j ,
k+1 (λq )i,1 y
−H
, k+1 (λq )i,N J −1 y
.
A.4 Representer calculation
209
k vˆ 0, ˆ kN I, j , j = v k vˆ kN I +1, j = vˆ 1, j, k ˆ Nk I, j , qˆ 0, j = q k qˆ Nk I +1, j = qˆ 1, j.
A.4
Representer calculation
A.4.1
Preamble
The ﬁrst part of the exercise is to calculate some representers, assemble the representer matrix and verify its algebraic properties. The second part is to ﬁnd the extremum of a penalty functional by solving the EL equations.
A.4.2
Representer vector
Calculate the representer vector r(x, y, t). Source code: rep.f To compile the source code: f 77O3 rep.f o rep
A.4.3
Representer matrix
Construct the representer matrix R. Check: Is R symmetric and positivedeﬁnite?
A.4.4
Extremum
Find the extremum of the penalty functional J . Note: Avoid storing r(x, y, t) when assembling (3.24). Instead, substitute for the coupling, integrate the backward EL equations and then integrate the forward equations (see §3.1: Accelerating the representer calculation). ˆ m , ym , tm ) − dm ]? Check: Is βˆ m = −w [q(x
A.4.5
Weights
The following values are suggested: Dynamical weight (u) W uf xyt = (0.25Fu )−2 s4 m−2 .
210
Appendix A Computing exercises
Dynamical weight (v) W vf xyt = (0.25Fu )−2 s4 m−2 . Dynamical weight (q) q
W f xyt = ∞. Initial weight (u) Wiu xy = ∞. Initial weight (v) Wiv xy = ∞. Initial weight (q) q
Wi xy = ∞. Boundary weight (v) Wbv xt = ∞. Data weight (q) w = [0.1 max(q F )]−2 m−2 .
A.5
More representer calculations
A.5.1
Pseudocode, preconditioned conjugate gradient solver
Find the local extremum of the penalty functional J iteratively (see §3.1, Accelerating the representer calculation). Below is the “pseudocode” of a preconditioned conjugate gradient method (Golub & Van Loan, 1989) for solving U−1/2 (R + w −1 I)U−1/2 U1/2 βˆ = U−1/2 h, where U is a preconditioner. l = 0; βˆ 0 = 0; e0 = h; ω0 = ! e0 !2 /! h !2 while ωl < ωsc solve zl = U−1 el l =l +1 if l = 1 p1 = z 0 else γl = eT l−1 zl−1 /eT l−2 z l−2 pl = zl−1 + γl pl−1 endif αl = eT l−1 zl−1 /pT l (R + w−1 I)pl βˆ l = βˆl−1 + αl pl
A.5 More representer calculations
211
el = el−1 − αl (R + w −1 I)pl ωl = ! el !2 /! h !2 endwhile βˆ = βˆ l Source code: cgm.f (conjugate gradient solver). A.5.2
Convolutions for covarying errors
Use “nondiagonal” covariances for initial and dynamical errors (see §2.6, Smoothing norms, covariances and convolutions).
Appendix B Euler–Lagrange equations for a numerical weather prediction model
The dynamics are those of the standard σ coordinate, PrimitiveEquation model of a moist atmosphere on the sphere (Haltiner and Williams, 1980, p. 17). A penalty functional and the associated Euler–Lagrange equations are given in continuous form; CMFortran code for ﬁnitedifference forms is available at an anonymous ftp site. Details of the measurement functionals for reprocessed cloudtrack wind observations (see §5.4), and the associated impulses in the adjoint equations, have been suppressed here. The details may be found in the code.
B.1
Symbols
a0 , " λ, φ, σ, t u, v, σ˙ f ≡ 2" sin φ #, T, Tv , p, p∗ = p/σ q, ρ l ≡ ln p∗ Rd , Cpd Rv , Cpv = Rv /Rd , δ = Cpv /Cpd
212
earth’s radius, rotation rate longitude, latitude, sigma, time zonal, meridional, vertical velocity components Coriolis parameter geopotential, temperature, virtual temperature, pressure, surface pressure speciﬁc humidity, density log surface pressure gas constant, speciﬁc heat at constant pressure (both for dry air) gas constant, speciﬁc heat at constant pressure (both for water vapor)
B.2 Primitive Equations and penalty functional
Su , Sv , ST , Sq , Sl u ≡
!1
udσ, u ≡ u − u
0
ρu , ρv , ρT , ρq , ρl , ρσ Wu , WT , Wq , W# , Wl , Wσ Qu , Q T , Q q , Q # , Q l , Q σ µ, θ, κ, χ, ξ, ω J = J [u, σ˙ , #, T, q, l] F ≡ ∂ T /∂ Tv = 1 + ( −1 − 1)q D≡
prior estimates of sources of zonal momentum, meridional momentum, heat, humidity, mass vertical average, ﬂuctuation dynamical residuals, or errors in prior estimates of sources weights for residuals (inverses of prior estimates of covariances of dynamical residuals) prior estimates of covariances of dynamical residuals weighted residuals, or adjoint variables penalty functional, or estimator for residuals moisture factor
1 ∂ σ˙ σ˙ (v cos θ)θ uλ + + − a0 cos θ a0 cos θ ∂σ σ
[δu], [δv], [δ#], [δT ], [δ σ˙ ], [δl], [δq] uI , TI , qI , lI Vu , VT , Vq , Vl Ou , OT , Oq , Ol #∗ , V∗ , O∗ •, ◦ run , rvn , r Tn , r#n , rqn , rln , rσn˙ aun , avn , aTn , aφn , aqn , aln , aσn˙
B.2
213
divergence Euler–Lagrange equations for arbitrary variations of u, v, #, T, σ˙ , l, q prior estimates of initial values for u, T, q, l weights for residuals or errors in prior estimates of initial values (inverses of initial error covariances) prior estimates of initial error covariances for u, T, q, l orography, weight, error covariance fourdimensional and threedimensional inner products representers for nth iterate of Euler–Lagrange equations adjoint representers
Primitive Equations and penalty functional u v u u λ + u φ + σ˙ u σ − f + ut + tan φ v a0 cos φ a0 a0 1 + (#λ + Rd Tvlλ ) − Su ≡ ρu a0 cos φ u v u vt + vλ + vφ + σ˙ vσ + f + tan φ u a0 cos φ a0 a0 1 + (#φ + Rd Tvlφ ) − Sv ≡ ρv a0
(B.2.1)
(B.2.2)
214
Appendix B Numerical weather prediction
u v Tλ + Tφ + σ˙ Tσ a0 cos φ a0
Rd Tv (v cos φ)φ 1 σ˙ + uλ + +σ − ST ≡ ρ T Cpd B a0 cos φ a0 cos φ σ σ
Tt +
qt +
u v qλ + qφ + σ˙ qσ − Sq ≡ ρq a0 cos φ a0
∂# + Rd Tv ≡ ρ# ∂lnσ
u v 1 (v cos φ)φ lt + lλ + uλ + −Sl ≡ ρl lφ + a0 cos φ a0 a0 cos φ a0 cos φ
u v 1 (v cos φ)φ lλ + lφ + u λ + + σ˙ σ − Sl ≡ ρσ a0 cos φ a0 a0 cos φ a0 cos φ Tv ≡ T [1 + (ε −1 − 1)q],
(B.2.3)
p = ρ Rd Tv ,
(B.2.4) (B.2.5) (B.2.6)
B ≡ 1 + (δ − 1)q.
(B.2.7)
Weighted residuals µ ≡ Wu • ρu ,
θ ≡ WT • ρT ,
κ ≡ Wq • ρq ,
(B.2.8)
χ = W # • ρ# ,
ξ = Wl • ρl ,
ω = Wσ • ρσ .
(B.2.9)
Penalty functional J = J [u, σ˙ , #, T, q, l] = ρ∗u • Wu • ρu + ρT • WT • ρT + ρq • Wq • ρq + ρ# • W# • ρ# + ρl • Wl • ρl + ρσ • Wσ • ρσ + (boundary penalties @ σ = 0, 1) + (initial penalties for u, T, q, & l) + (data penalties). (B.2.10)
B.3
Euler–Lagrange equations
Note: The symbols [δu], etc., indicate that the following equation is the extremal condition for the penalty functional J , with respect to variations of δu, etc. [δu] uλµ tan φ vλ ν (uµ)λ (vµ cos φ)φ vµ + − − − (σ˙ µ)σ − a0 cos φ a0 cos φ a0 cos φ a0 a0 cos φ tan φ u tan φ ν+ + f + uν a0 a0 Tλ θ 1 Rd Tv θ qλ κ lλ ξ ξλ + − + + − a0 cos φ Cpd B λ a0 cos φ a0 cos φ a0 cos φ a0 cos φ
−µt +
+
lλ (ω − ω) (ωλ − ωλ ) − − (impulses) = 0 a0 cos φ a0 cos φ
(B.3.1)
B.3 Euler–Lagrange equations
[δv]
215
vφ ν (uν)λ u tan φ (vν cos φ)φ + − (σ˙ ν)σ µ− f + − a0 a0 cos φ a0 a0 cos φ 1 Tφ θ Rd Tv θ qφ κ lφ ξ ξφ lφ (ω − ω) + − + + − + a0 Cpd B φ a0 a0 a0 a0 a0
uφ µ −νt + − a0
−
(ωφ − ωφ ) − (impulses) = 0 a0
(B.3.2)
[δ#] −
µλ (ν cos φ)φ − − (σ χ )σ − (impulses) = 0 a0 cos φ a0 cos φ
(B.3.3)
[δT ] −θt +
Rd Flλ µ (uθ )λ Rd FDθ Rd Flφ ν (vθ cos φ)φ − + − − (σ˙ θ)σ + a0 cos φ a0 a0 cos φ a0 cos φ Cpd B
+ Rd Fχ − (impulses) = 0
(B.3.4)
[δ σ˙ ] Rd u σ µ + vσ ν + Tσ θ − Cpd
Tv θ B
σ
−
Rd Tv θ + qσ κ −ωσ − (impulses) = 0 Cpd Bσ (B.3.5)
[δl] −ξt − −
Rd (Tv ν cos φ)φ (uξ )λ (vξ cos φ)φ Rd (Tv µ)λ − − − a0 cos φ a0 cos φ a0 cos φ a0 cos φ u ωλ v ω cos φφ − − (impulses) = 0 a0 cos φ a0 cos φ
(B.3.6)
[δq] Rd TFq lλ µ Rd T Dθ Rd TFq lφ ν Rd Dθ (δ − 1)Tv + Fq + − a0 cos φ a0 Cpd B Cpd B2 (vκ cos φ)φ (uκ)λ − + Rd Fq χ −(σ˙ κ)σ −(impulses) = 0 − a0 cos φ a0 cos φ
−κt +
(B.3.7)
Note: (B.3.5) is a ﬁrstorder, ordinary differential equation for ω as a function of σ . Solutions are indeterminate, since there is no boundary condition for ω at σ = 0, nor at σ = 1. However, the other Euler–Lagrange equations only involve ω or u ω, where ω ≡ ω − ω, etc., thus the indeterminacy has no effect upon them. The residual ρσ in (B.2.6) must satisfy ρσ = 0; thus its covariance Q σ must have vanishing integrals with respect to both vertical
Appendix B Numerical weather prediction
216
arguments, and any hypothesis for Q σ must conform to this requirement. Hence the optimal estimate ρσ ≡ Q σ • ω is also unaffected by the indeterminacy, and the vertical integral of the estimate vanishes. Initial conditions @ t = 0:
u∼ = uI ,
T ∼ = TI ,
q∼ = qI ,
l∼ = lI .
(B.3.8)
Contribution to penalty functional JI = (u − uI )∗ ◦ Vu ◦ (u − uI ) + (T − TI ) ◦ VT ◦ (T − TI ) + (q − qI ) ◦ Vq ◦ (q − qI ) + (l − lI ) ◦ Vl ◦ (l − lI ).
(B.3.9)
Hence @ t = T: @t = 0:
µ = 0,
θ = 0,
κ = 0,
−µ + Vu ◦ (u − uI ) = 0 , −κ + Vq ◦ (q − qI ) = 0,
i.e., u = uI + Ou ◦ µ,
T = TI + OT ◦ θ,
ξ =0
(B.3.10)
−θ + VT ◦ (T − TI ) = 0, −ξ + Vl ◦ (l − lI ) = 0, q = qI + Oq ◦ κ,
where Ou (12) ◦ Vu (23) = δ(x1 − x3 )I, etc.
(B.3.11)
l = lI + Ol ◦ ξ, (B.3.12)
Boundary conditions @ σ = 0, 1: @ σ = 1:
σ˙ = 0 #∼ = #∗ .
(B.3.13) (B.3.14)
Contribution to penalty functional J∗ = (# − #∗ ) ◦ V∗ ◦ (# − #∗ ).
(B.3.15)
Hence
B.4
@ σ = 0:
χ = 0,
(B.3.16)
@ σ = 1:
# = #∗ − O∗ ◦ χ .
(B.3.17)
Linearized Primitive Equations
Note: The labels (LPE1) etc. refer to lines of the CMFortran code for the ﬁnitedifference equations; the code is available at an anonymous ftp site (ftp.oce.orst.edu, /dist/chua/IOM/IOSU). u nt
v n−1 u nφ u n−1 u nλ u n−1 tan φ n−1 n vn + + + σ˙ u σ − f + a0 cos φ a0 a0 1 n #λ + Rd Tvn−1lλn − Sun ≡ ρun . + (LPE1) (B.4.1) a0 cos φ
B.5 Linearized Euler–Lagrange equations
vtn
217
v n−1 vφn u n−1 vλn u n−1 tan φ n−1 n un + + σ˙ vσ + f + + a0 cos φ a0 a0 1 n + (LPE2) (B.4.2) # + Rd Tvn−1lφn − Svn ≡ ρvn . a0 φ
v n−1 Tφn u n−1 Tλn + + σ˙ n Tσn−1 a0 cos φ a0 n % & vφ cos φ φ u nλ ∂ σ˙ n σ˙ n Rd Tvn−1 + + − − STn ≡ ρTn . + Cpd B n−1 a0 cos φ a0 cos φ ∂σ σ (LPE3) (B.4.3)
Ttn +
qtn +
ltn +
v n−1 qφn u n−1 qλn + σ˙ n−1 qσn − Sqn ≡ ρqn . + a0 cos φ a0
(LPE4) (B.4.4)
∂#n + Rd T n F n−1 ≡ ρ#n . ∂lnσ
(LPE5) (B.4.5)
/ 0 v n−1 lφn u n−1 lλn 1 (v n cos φ)φ + + u n λ + − Sln ≡ ρln . a0 cos φ a0 a0 cos φ a0 cos φ (LPE6) (B.4.6)
v n−1lφn u n−1lλn ∂ σ˙ n 1 (v n cos φ)φ + u n + − Sln−1 ≡ ρσn˙ . + + λ a0 cos φ a0 a0 cos φ a0 cos φ ∂σ (LPE7) (B.4.7) Tvn = T n [1 + (ε −1 − 1)q n ]. (B.4.8)
B.5 [δu n ]
Linearized Euler–Lagrange equations (u n−1 µn )λ u n−1 tan φ (v n−1 µn cos φ)φ νn − − (σ˙ n−1 µn )σ + f + a0 cos φ a0 cos φ a0 n n−1 n ωλ − ωn λ Tv θ 1 Rd ξλn − − − Cpd a0 cos φ B n−1 λ a0 cos φ a0 cos φ
− µnt −
n−1 u n−1 v n−1 tan φ n−1 vλn−1 ν n−1 tan φ n−1 n−1 λ µ µ − u ν + − a0 cos φ a0 a0 cos φ a0 q n−1 κ n−1 lλn−1 ξ n−1 T n−1 θ n−1 − λ − − λ a0 cos φ a0 cos φ a0 cos φ n−1 n−1 (ω − ω ) − lλn−1 + (impulses)n . (LELE1) (B.5.1) a0 cos φ
=−
Appendix B Numerical weather prediction
218
[δv n ] (u n−1 ν n )λ (v n−1 ν n cos φ)φ − − (σ˙ n−1 ν n )σ − a0 cos φ a0 cos φ n / n 0 ξφn ωφ − ωφ Rd Tvn−1 θ n 1 − − − Cpd B n−1 φ a0 a0 a0
− νtn −
f +
u n−1 tan φ a0
µn
n−1 u n−1 vφn−1 ν n−1 qφn−1 κ n−1 lφn−1 ξ n−1 Tφn−1 θ n−1 φ µ − − − − a0 a0 cos φ a0 a0 a0 n−1 lφ − (ωn−1 − ωn−1 ) + (impulses)n . (LELE2) (B.5.2) a0
=−
[δ#n ] −µnλ (ν n cos φ)λ − − (σ χ )nσ = (impulses)n . a0 cos φ a0 cos φ
(LELE3) (B.5.3)
[δT n ] (u n−1 θ n )λ (v n−1 θ n cos φ)φ − + Rd F n−1 χ n a0 cos φ a0 cos φ Rd F n−1lφn−1 ν n−1 Rd F n−1lλn−1 µn−1 Rd F n−1 Dn−1 θ n−1 − =− − a0 cos φ a0 cpd B n−1 n−1 n−1 + σ˙ θ + (impulses)n . (LELE4) (B.5.4) σ
−θnt −
[δ σ˙ n ] Tσn−1 θ n −
Rd Cpd
Tvn−1 θ n B n−1
σ
−
Rd Tvn−1 θ n − ωσn Cpd B n−1 σ
n−1 = −u n−1 − vσn−1 ν n−1 − qσn−1 κ n−1 + (impulses)n . σ µ
[δl n ] −ξtn
− −
Rd
/
Tvn−1 µn a0 cos φ
0 λ
−
Rd
/
Tvn−1 ν n cos φ
0
a0 cos φ
−
(u n−1 ξ n )λ a0 cos φ
(v n−1 ξ n cos φ)φ (u n−1 ωn )λ (v n−1 ωn cos φ)φ − − a0 cos φ a0 cos φ a0 cos φ
= (impulses)n .
[δq n ] −κtn −
φ
(LELE5) (B.5.5)
(u n−1 κ n )λ (v n−1 κ n cos φ)φ − − (σ˙ n−1 κ n )σ a0 cos φ a0 cos φ
(LELE6) (B.5.6)
B.6 Representer equations
=− +
Rd T n−1 Fqn−1lλn−1 µn−1 a0 cos φ
Rd T n−1 Fqn−1lφn−1 ν n−1
−
a0
−
219
Rd T n−1 Fqn−1 Dn−1 θ n−1 Cpd B n−1
Rd Tvn−1 Dn−1 θ n−1 (δ − 1) − Rd Fqn−1 χ n−1 + (impulses)n . Cpd B 2
(LELE7) (B.5.7)
Representer equations
B.6 [u]
runt +
u n−1runλ
+
v n−1runφ
+ σ˙ n−1runσ −
f +
u n−1 tan φ a0
rvn
a0 cos φ a0 1 n + r#λ + Rd Tvn−1rlnλ = Qu • anu u . a0 cos φ
(RE1) (B.6.1)
[v] rvnt
v n−1rvnφ
u n−1rvnλ
[T ] r Tnt
v n−1r Tnφ
u n−1r Tnλ
+ a0 cos φ a0 n rσ˙ +σ = Q T • aTn . σ σ
+
u n−1 tan φ f + a0
+ + + a0 cos φ a0 1 n + r#φ + Rd Tvn−1rlnφ = Qu • anu v . a0 +
σ˙ n−1rvnσ
+ rσn˙ Tσn−1
run (RE2) (B.6.2)
cos φrvn φ runλ Rd Tvn−1 + + Cpd B n−1 a0 cos φ a0 cos φ (RE3) (B.6.3)
[q] rqnt +
u n−1rqnλ
a0 cos φ
+
v n−1rqnφ a0
+ σ˙ n−1rqnσ = Q q • aqn .
(RE4) (B.6.4)
[#] ∂r#n n . + Rdr Tn F n−1 = Q # • a# ∂lnσ [l] rlnt
+
u n−1 rlnλ a0 cos φ
+
v n−1 rlnφ a0
/ 0
/ n0 cos φ rvn φ ru λ + + = Q l • aln . a0 cos φ a0 cos φ (RE6) (B.6.6)
[σ˙ ] u n−1rlnλ
a0 cos φ
+
v n−1rlnφ a0
+
(RE5) (B.6.5)
runλ
a0 cos φ
+
n rv cos φ φ a0 cos φ
+
∂rσn˙ = Q σ˙ • aσn˙ . ∂σ (RE7) (B.6.7)
Appendix B Numerical weather prediction
220
Representer adjoint equations
B.7 [δu]
n−1 n n−1 n v a cos φ n−1 n a u u n−1 tan φ u φ u n λ − − σ˙ au σ + f + avn − au t − a0 cos φ a0 cos φ a0 / 0 n aσ˙λ − aσn˙λ alnλ 1 Rd Tvn−1 aTn − − − = (impulse). Cpd B n−1 λ a0 cos φ a0 cos φ a0 cos φ
(RAE1) (B.7.1) [δv]
n−1 n n−1 n v av cos φ φ n−1 n a u u n−1 tan φ v λ n −avt − aun − − σ˙ av σ − f + a0 cos φ a0 cos φ a0 / 0 n aσ˙φ − aσn˙φ alnφ Rd 1 Tvn−1 aTn − − − = (impulse). Cpd a0 B n−1 φ a0 a0 (RAE2) (B.7.2)
[δ#] −
aunλ
a0 cos φ
−
n av cos φ φ a0 cos φ
n = (impulse). − σ a# σ (RAE3) (B.7.3)
[δT ]
n−1 n v aT cos φ φ u n−1 aTn λ n − = (impulse). − + Rd F n−1 a# a0 cos φ a0 cos φ (RAE4) (B.7.4)
−aTnt [δ σ˙ ]
Tσn−1 aTn − [δl]
Rd Cpd
Tvn−1 aTn B n−1
σ
−
Rd Tvn−1 aTn ∂aσn˙ − = (impulse). Cpd B n−1 σ ∂σ (RAE5) (B.7.5)
0 / n−1 n n−1 n n−1 n 0 Tv av cos φ φ v al cos φ φ u al λ Tvn−1 aun λ n −Rd − Rd − alt − − a0 cos φ a0 cos φ a0 cos φ a0 cos φ / n−1 n 0 / n−1 n 0 v aσ˙ cos φ φ aσ˙ λ u − − = (impulse). (RAE6) (B.7.6) a0 cos φ a0 cos φ /
[δq]
−aqnt −
u n−1 aqn
a0 cos φ
λ
−
v n−1 aqn cos φ a0 cos φ
φ
− σ˙ n−1 aqn σ = (impulse). (RAE7) (B.7.7)
Author index
Adams, R. A., 55 Adcroft, A., 167 Ahlquist, J., 84 Amodei, L., 62 Andersen, O. B., 136 Anderson, D. L. T., xx Andrews, P. L. F., 166 Arakawa, A., 197 Azencott, R., 116
Ballard, S. P., 166 Baraille, R., 104, 105 Barker, D. M., 166 Barker, E., 165, 166 Barth, N., 116 Batchelor, G. K., xix Baugh, J. R., 61 Bell, R. S., 166 Bennett, A. F., xx, 17, 23, 32, 38, 44, 45, 49, 51, 58, 61–64, 67, 70, 72, 76, 84, 103, 131, 135, 136, 139, 148, 149, 151–153, 156–161, 163, 164, 168, 169, 171, 173, 180, 182, 183, 185 Bleck, R., 79 Bogden, P., 168 Bray, J. R., 166 Bryan, K., 83 Bretherton, F. P., 38 Busalacchi, A. J., 51, 156
Callahan, P. S., 124 Canceil, P., 133, 136 Cane, M. A., 5, 79, 83, 144, 156 Cao, X., 167 Carton, J. A., 167 Carton, X., 104, 105 Cartwright, D. E., 125, 136 Chan, N. H., 104 Chandra, R., 58 Chapman, S., 150 Chelton, D. B., xxi, 124 Chepurin, G., 167 Chua, B. S., xxi, 51, 63, 72, 148, 151–153, 156–161, 163, 164, 171, 180, 182, 183, 185 Clayton, A. M., 166 Cohn, S. E., 104 Courant, R., 8–10, 15, 173 Courtier, P., xx, 51, 62, 170, 171 Cowling, T. G., 150 Cox, D. R., 109 Cox, M. D., 83 Dagum, L., 58 Dalby, T., 166 Daley, R. A., 38, 51, 70, 165, 166 Da Silva, A. M., 167 Davidson, N. E., 148 Davis, R. E., 38, 167, 168 Dee, D. P., 167
221
222
Author index
De Mey, P., 104, 105 Denbo, D. W., 156 Derber, J. C., xx, 51, 64, 90, 167 Doodson, A. T., 125 Dutz, J., 1–3 Egbert, G. D., xxi, 62–64, 109, 131, 135–137, 165 Eknes, M., 169 Elsberry, R. L., 148 Emanuel, K., 148 Entekhabi, D., 169 Errico, R. M., xx, 51, 93 Evensen, G., 169 Fandry, C. B., 38 Ferziger, J. H., 145 Flannery, B. P., xix, 43, 51, 89, 166 Flather, R. A., 136 Fomin, S. V., 94, 97 Foreman, M. G. G., 62, 64, 131, 135, 136 Francis, O., 136 Franklin, J. L., 151 Fu, L.L., 104, 124, 167, 168 Fukumori, I., xx, 104, 105, 143, 167, 168 Gao, J., 84 Gelb, A., 98 Gelfand, I. M., 94, 97 Genco, M. L., 133, 136 Gent, P. R., 79 Ghil, M., xx, 170, 171 Giering, R., 167, 168 Gill, A. E., xix, 118, 124 Goerss, J. S., 166 Golub, G. H., 210 Goodrich, R. K., 167 Greiner, E., 90 Hackert, E. C., 51, 156 Hadamard, J., 8 Hagelberg, C. R., 149 Haines, K., xx Haltiner, G. J., xix, 11, 118, 151, 212 Hanes, B. J., 124 Harrison, D. E., 51, 72, 156–161, 163, 164 Hartley, D. E., xx Hayden, C. M., 151 Heimann, M., xx Heisey, C., 167
Hilbert, D., 8–10, 15, 173 Hill, C., 167 Hinkley, D. V., 109 Hoang, S., 104, 105 Hobbs, P. V., 150 Hodur, R. M., 166 Hogan, T. F., 166 Holland, G. J., 148 Holton, J., xix Ide, K., xx, 170, 171 Ingleby, N. B., 166 J¨arvinen, H., 167 Jenne, R. L., 156 Kadane, J. B., 104 Kamachi, M., 170 Kaminsky, T., 167 Kasibhatla, P., xx Kelly, G., 167 Keppenne, C. L., 169 Kessler, W. S., 156–158 Kimoto, M., xx King, C., 136 Kleeman, R. C., 90, 156 Klinker, E., 167 Kloeden, P. E., 173 Klosko, S., 136 Kohr, D., 58 Kruger, J., 116 Kundu, P. K., xix Ladyzhenskaya, O. A., 173 Lagarde, T., 171 Lanczos, C., 15 Le Dimet, F.X., 77, 90, 92 Le Provost, C. L., 133, 136, 137 Lee, T., 167, 168 Leslie, L. M., 63, 149, 151–153 Lewis, J. M., 90 Li, D., 166 Li, X., 136 Lions, J. L., 77 Ljung, L., xv Long, R. B., 90 Lorenc, A. C., 111, 166, 170, 171 Louis, J.F., xx, 51 Luong, B., 92 Lyard, F. H., 133, 136, 137 Lynch, J. S., 151
Author index
Mahfouf, J.F., 167 Mahowald, N., xx MalanotteRizzoli, P., xx, 64, 105 Marotske, J., 167, 168 Marshall, J., 167, 168 Maydan, D., 58 McAvaney, B. J., 148 McClurg, D. C., 156 McCreary, J. P. Jr., 156 McDonald, J., 58 McIntosh, P. C., 58 McLaughlin, D. B., 169 McPhaden, M. J., 51, 72, 156–161, 163, 164 Meditch, J. S., 94 Menemenlis, D., 167, 168 Menon, R., 58 Menzel, W. P., 151 Mesinger, F., 197 Metropolis, N., 112 Miller, R. N., xxi, 17, 51, 84, 104, 156 Molines, J. M., 136 Moore, A. M., 90, 156 Muccino, J. C., 93, 169 Nagata, M., xx Navon, I. M., 84 Ngodock, H.E., xxi, 77, 92 Niiler, P., 167, 168 Oliger, J., 172, 173, 183 Pacanowski, R. C., 80 Pacheco, P., 58 Palma, W., 104 Parke, M., 136 Parker, R. L., xvi Passi, R. M., 167 Payne, T. J., 166 Pedder, M. A., 38 Pedlosky, J., xix Perelman, L., 167 Perigaud, C., 90 Philander, S. G. H., 80 Phoebus, P. A., 166 Piacentini, A., 171 Powers, P. E., 149 Press, W. H., xix, 43, 51, 89, 166 Prinn, R. G., xx Rabier, F., 167 Ray, R. D., 136, 137
Rayner, P., xx Reichle, R. H., 169 Reid, W. T., xv Renton, M. W., 156 Reynolds, R. W., 156 Ries, J. C., 124 Rodgers, C. D., xvi Rosati, A., 64 Rosenbluth, A. W., 112 Rosenbluth, M. N., 112 Rosmond, T. E., 166 Saiki, M., xx Sasaki, Y., 25, 167 Sato, N., xx Saunders, F. W., 166 Schiff, L. I., 12 Schrama, E. J. O., 136 Sheinbaum, J., xx Shum, C. K., 136 Simmons, A., 167 Slutz, R. J., 156 Smith, L. T., 79 Smith, N. R., 90, 156 Smith, T. M., 156 S¨oderstr¨om, T., xv Soreide, N. N., 156 Spillane, M. C., 157, 158 Stakgold, I., xix, 179 Stammer, D., 167, 168 Steurer, P., 156 Sundstr¨om, A., 172, 173, 183 Sweeney, D., cover photograph Talagrand, O., 77, 90, 104, 105 Teller, A. H., 112 Teller, E., 112 Teukolsky, S. A., xix, 43, 51, 89, 166 Thacker, W. C., 90 Thi´ebaux, H. J., 38 Thorburn, M. A., 76, 139 Thual, O., 171 Uboldi, F., 170 Van Loan, C. F., 210 Velden, C. S., 151 Verron, J., 92 Vetterling, W. T., xix, 43, 51, 89, 166
223
224
Author index
Vincent, P., 133, 136 Vukicevic, T., xx, 51
Xu, L., 70 Xu, Q., 72, 84
Wahba, G., xvii, 33, 53 Wallace, J. M., 150 Warburg, H. D., 125 Weinert, H. L., 38 Wendelberger, J., xvii, 33, 53 Williams, R. T., xix, 11, 118, 151, 212 Woodruff, S. D., 156 Woodworth, P. L., 136 Wunsch, C., xx, 116, 136, 137, 167, 168
Yoshida, K., 32 Young, R. E., 104 Zaron, E. D., 84 Zebiak, S. E., 5, 83, 144, 156 Zhang, K. Q., 167 Zhang, S., 84 Zhu, W. H., 156 Zlotnicki, V., 167, 168 Zou, X., 84
Subject index
20◦ isotherm (Z20), 5, 157–165 3DVAR, 166 4DVAR, 167 βplane, 157 f plane, 126 F test, 3 ndimensional space, 55 rk (reproducing kernel), 23, 29, 37, 38, 47 χ 2 , xv, 6, 43
abstraction, xvii abstraction, na¨ıve, 146 Acartia clausi, 1 acoustic path, 45 tomography, reciprocalshooting, 128 transceiver, 116 adjoint, 19 discrete equation, 93 dynamics, 76–77 dynamics, destabilized, 77 equation, 19, 30 ﬁrstguess, 77 operator, xvii, 77, 134 representer, 22, 29 symmetry, 93 variable, 28, 98 advection, 132, 141 anomalous, 156 nondivergent, 75
advective coupling, 77 Alexandrium lusitanicum, 1 algebraic power, 55 algebraic system, 50 algorithm cycling, 103 descent, 91 direct representer, 59, 67, 103 explicit, 97 indirect representer, 61, 63, 72, 78, 135 iterated indirect representer, 6, 146, 160, 169, 183, 216–220 open loop, 59, 60, 63 productpolynomial, 167 Rauch, Tung, and Streibel, 98 substituting, 113 sweep, 94 symbolic, 167 aliasing, 143 altimetric missions, GEOSAT, ERS1, TOPEX–POSEIDON, 128 amplitude modulation, 74, 77 amplitude perturbation, 79 analysis, xiii, xv, 40, 41, 104, 111, 148, 149, 151, 166 array mode, 135 optimal interpolation, 148 annealing, simulated, xviii, 85, 115, 116, 182, 183 anomalies of tropical ﬁelds, 156 Antarctic Circumpolar Current, 179 Arakawa Cgrid, 197 arithmetic mean, 3, 106, 108
225
226
Subject index
array, 48, 50 assessment, 5, 51, 135 mode, 49–51 Australian Bureau of Meteorology Research Centre (BMRC), 148 average arithmetic, 3, 106, 108 ensemble, 145 Reynolds, 41 sample, 71 space, 145 subgridscale, 41 time, 145 background, 11, 111, 166, 170 backward equation, 19 balance, condition of, 112 balance, condition of detailed, 112, 113 balloon, high altitude, 142 bare impulse, 19 basin resonance, 133 bathymetry, 130, 133 Bayes’ Rule, 109, 111 bellshaped correlation, 54, 64, 158, 167 best linear unbiased estimate (BLUE), 39, 40, 72, 99 bestﬁt, xv, 36 beta plane, equatorial, 157 bias, 39, 41, 72, 74, 99, 101, 108, 128, 167 biconjugategradient, 93 bilinear function, 52 bilinear interpolation, 156 bimodal distribution, 105 biological model, 116 biology, marine, xix boundary condition, xiv, 8, 9, 27, 50, 52, 127, 130, 134 computational, 180 free slip, 157 heatreservoir, 94 meridional, 157 no slip, 157 partitioning of, 183 periodic, 198, 200 rigid, 133, 198, 200 zonal, 157 boundary error, uncorrelated in time, 96 layer, 103 open, 127, 180 operator, 21, 22 penalty, 18 boundary value error covariance for inverse estimate, 67 excess, xx
problem, xiv, xix twopoint problem in time, 19, 66 bruteforce minimization, xviii, 85 calculus of variations, xvi, 15, 52, 56, 83, 85, 134 calculus of variations, secondorder information, 92 Cane–Zebiak coupled model, 144 chain rule, 120 chaotic solution, 12 checkerboarding, 115 checklist, data assimilation, xix, 4 chisquared (χ 2 ) test, xv, 6, 43 Cholesky factorization, 166 climate prediction, seasonaltointerannual, 156 climatology plus persistence (CLIPER), 150 clipping, 50 cluster point, 142 coding error, 93 coefﬁcient matrix, 50 color, 72 completing the square, 34 compromise, 104, 167 computational boundary condition, 180 degrees of freedom, xviii efﬁciency, 18 harm, 84 practice, 72 requirements, 74 computer architecture, multiprocessor, xviii, 61 conditioning, xv, 4, 51 conditioning, discrete analysis of, 51 conduction, 93 conjugate gradient, preconditioned, 146, 166 conservation entropy, 121, 144 mass, 144 momentum, xvi, 121 scalar, 121 total vorticity, 141 volume, 79, 119 continuity, 56 continuous dependence, 8, 173 random variable, 105 solution, 139 continuum, xx, 50, 51 control space, xvii, xviii, 51 convection, xviii, 83, 84, 116 convection, timing penalty, 84 convergence, 62, 63, 74, 76, 92, 142, 169 convolution, 29, 51, 60, 64, 65 coordinate isobaric, 118, 121, 177
Subject index
isopycnal, 177 sigma, 151, 212 Coordinated Universal Time (UTC), 152 Coriolis parameter, 114, 121, 126, 142, 212 correlation anisotropic, 158 bellshaped, 54, 64, 158, 167 Markovian, 158 cost functional, 14 Coupled Ocean Atmosphere Meteorological Prediction System (COAMPS), 166 coupling vector, 62 Courant–Friedrich–Lewy stability criterion, 11 covariance, xv, xix, 6, 37, 38, 40, 48, 53 bellshaped, 54, 64, 158, 167 cross, 99, 168 effectively hypothesized, 56 equilibrium, 105 “explained”, 71 hypothesized, 51, 56, 68 inhomogeneous, 65 inverse, 68 Markovian, 158 matrix, for data error, 40, 48, 62 posterior error, 66, 67, 70, 72, 74 prior error, 37, 48, 51, 67, 71 sample, 64 guided by model dynamics, 146 harmonic mean of two, 57 nonintuitive structures, 155 nondiagonal, 64 crossover difference, 128, 131, 134 current meter, 45 cycling, algorithm, 103 cycling, inverse, 103 cycling, sequence, 142 data assimilation checklist, xix derived, 149 error, 24 exchange of, 61 impact, 103 impulse, 76 large set of, 59 misﬁt, 22, 44, 73 penalty, 17, 18, 75 perfect, 24 pointwise, 13, 40 prescribed continuously on curves, 181 projection, 49 reduce error variance, 102 status, 59 subspace, xviii, 5, 51, 74, 78 sufﬁciently dense, 71 synthetic, 183 vector, 21 weight, 22, 23, 40 worthless, 24
227
decomposition, singular value, 5 degrees of freedom, xv, 3, 36, 43, 105, 136, 149, 160, 170 computational, xviii, 116 effective, 154 observable, 36 delta, Dirac, 10, 16, 18, 52 delta, Kronecker, 18, 21 descent algorithm, 91 determinism, 8 diagnostic constraint, xvi diagnostic equation, 114 differentiability, 8, 56 differentiable function, 8 differential order, 55 diffusion, 27, 51 direct solution, 61 discontinuity, smooth over, 84 discrete analog, 78 discrete equation, adjoint, 93 dispersion, 27, 169 divergence, 141 documentation, 170 domain, xiv bounded, 53, 55 comoving, 185 doublyperiodic, 142 entire plane, 53 morphology, 146 planar, 52 simplyconnected, 123, 138 spatially ﬁnite, 65 dominance, diagonal, 50 downhill strategy, 115 drag, 126, 133 dynamics, 5, 22 adjoint, 169 error in, xvii linearized, 169 nonlinear, 24 nonsmoothness of, 74 strong constraint, 87, 160 unrealistic, 80 weak constraint, 6 eddyﬂux, 145 Egbert’s Table, 109, 110 eigenmode, 179 eigenvalue, 49–51, 175, 179 eigenvector, 49, 50, 176 Ekman layer dynamics, 169 El Ni˜no Southern Oscillation (ENSO), xx, 5, 156, 158 elevation, sealevel, 45 empirical orthogonal function (EOF), 149
228
Subject index
encapsulation of prior knowledge, 48 energy, total mechanical, 184 energy integral, 11, 174, 175, 184, 185 ensemble, 37, 41, 113, 145 ensemble average, 37, 39, 41, 160 Ensemble Kalman Filter (EnKF), xvi, 169 equation abstract ﬁnitedimensional, 170 elliptic, 114 ﬁrstorder wave, 18, 114 forward, 30 homogeneous, 52 Korteveg–DeVries, 169 momentum, linearized, 77, 79 nonlinear wave, 75, 81 normal, 3, 4 of motion, 200 of state, 122 partial differential, xiv “poor man’s balance”, 142 sequences of, 78 shallowwater, xix, 5, 122, 125, 173, 180, 182 shallowwater, nonlinear, solution uniqueness, 183 equilibrium, chain in, 112 equilibrium covariance, 105 error bathymetric, 134 drag, 134 dynamical, 64 forcing, 41 intermittent, xviii isotherm depth, 157 model, 51 nonGaussian, xviii observing system, 51 of interpretation, 142 parameterization, 41 posterior statistics, 5 random, 37 random measurement, 42 SST, 157 underestimated, 41 wind component, 157 error covariance, 6, 102 posterior, 66, 67, 70, 72, 74 posterior, sample mean, 73 prior, data, 72 spatial, 100 estimate best linear unbiased (BLUE), 39, 40, 72, 99 data residual, 5 dynamical residual, 5 inverse, 36, 64 mathematical smoothness of, 55 posterior error, 3 prior 20, 36, 39, 81 regular, 29 sequential, 94
state, 5 state, error in, 69 statistical, 53 Estimation of the Circulation and Climate of the Ocean Consortium (ECCO), 167 estimator, xvii, 2, 5, 6, 68, 84, 168 ETOP95 bathymetry, 133 Euler–Lagrange equations, xvii, 6, 12, 16–19, 24, 26, 28, 29, 47, 50, 51, 55, 61, 73, 75–77, 83, 85, 90, 95, 140, 168, 181, 182, 202, 214 discrete, 90 illposed for backward integration, 80 linear, 59, 76, 217 our most general form, 95 nonlinearity, 82 practical experience of nonlinear iteration, 74 representer solution of, 29 steps for solving, 22 European Centre for Mediumrange Weather Forecasting (ECMWF), 151, 167 existence, 173 extremum, 16 fast talk, 145 ﬁddle, 84 ﬁlter, gappy running mean, 156 ﬁltering of data, 143, 156 ﬁnite difference, xx, 11, 17, 50, 72, 83, 88, 93, 136, 166 Finite Element Model (FEM), 133, 136 ﬁrstguess, xvi, 11, 38, 61, 136, 203 Fleet Numerical Meteorology and Oceanography Center (FNMOC), xxi, 166 ﬂoat, deeply submerged, 142 forcing, prior estimate of, 8 forecast, honest, 149 forecast skill, 149, 167 FORTRAN, xviii, xix, 146, 195 forward problem, 13 forward problem, overdetermined linear, 59 Four Dimensional Variational Assimilation (4DVAR), 167 Fourier transform, 53 Fourier transform, inverse, 54 free surface, 122 front, 183 Froude number, 185 functional, 45 analysis, 29, 88 differentiation, 78 evaluation, 47 linear, 33, 46 fudge, 75 fundamental solution, 10, 12
Subject index
garbage, 45 Gauss–Markov smoothing, xviii Gaussian, multivariate, 168 Gaussian, non, xviii Gaussian (normal) variable, xvii, 38 generalized inverse, 12, 21, 40, 41, 59, 61, 66 condition or stability of, 50 geometical structure, xvii, 32 geoid, 127, 128 geometrical interpretation, 32 geophysical ﬂuid dynamics, xviii, 74, 83 Geophysical Fluid Dynamics Laboratory (GFDL), 167 geophysical inverse theory, xvi geopotential, 120, 142, 158, 177 geostrophy, xvi, xix, 123, 142, 166 global minimum, 15 gradient, 78, 82, 87 information, 18, 85, 90, 91 na¨ıve evaluation, 87 preconditioned conjugate, 146, 166 search, xviii, 78, 82, 84 gravity wave, 79 Green’s function, xix, 8, 10, 12, 21–23, 26, 27, 29, 38, 47, 69, 100 grid, 144 Arakawa C, 197 coarse, 73, 104, 136, 151 ﬁne, 151, 160 spacing in shallow water, 133 Gulf Stream meanders, 143 halfpower point, 54 harmonic analysis, 52, 134 heat transfer, turbulent, 80 Heaviside unit step function, 10 Hessian, 78, 92, 93 Hilbert Space, xvii, 32, 33, 46 histogram, 113 Hybrid Monte Carlo (HMC), 182 hydrodynamics, xix hydrography, 24 hydrostatic approximation, xix, 119, 120, 126 hypothesis, xv, 5, 6, 25, 41, 51, 67, 68, 81, 82, 146, 216 null, 3, 37, 42–44, 71, 163 reasonable, 92 rejection, 51, 74 suspect, 41, 82 test, 42, 51 illconditioned, 4 illposed problem, 8, 12, 80, 172
importance sampling, 112, 183 impulse bare, 19 vector, 95 vector coefﬁcients, 60 incompressible ﬂow, 122, 173 inﬂow, 176 inﬂuence function, xix, 69 inhomogeneity, xiii, 8 initial condition, xiv, 8, 9, 27, 50, 66, 173, 200 error ﬁeld, sample, 72 operator, 21, 22 penalty, 17, 18 initial value boundaryvalue problem, mixed, xiv, 18 error covariances for inverse estimates, 67 problem, 19 varied independently, 91 inner product, 32, 33, 53 inputs, 4, 5 Institut de M´ecanique de Grenoble (IMG), 133, 136 integrals, convolutionlike, 51 interdisciplinary range, xx interference, constructive, 125 interference, destructive, 125 intermediate model, 5, 114, 156 intermittency, xviii, 105 interpolant, 39, 40, 101 interpolation bilinear, 156 in measurement, 63 limit of smoothing, 24 optimal, 40, 54, 111 univariate statistical, 148 Inverse Ocean Model (IOM), 171 isobaric coordinate, 118, 121, 177 isopycnal coordinate, 177 isotropic random ﬁeld, 53 iteration alternative scheme, 141 heuristic, 77 inner, 63 linearizing, 78 nonlinear leastsquares problem, 74 outer, 74, 78 parameter estimation, 169 Scheme A, 76 Scheme B, 77 simple, 81, 141 iterative solver, 61 Jacobian, 140 jet, xviii jet stream, 145
229
230
Subject index
Kalman ﬁlter, 156 as linear regression, 98 augmented, 104 continuity of, 97 controltheory, 94 ensemble, xvi, 169 equilibrium of error covariance, 102 estimate, 101 evolution time scale, 103 gain, 101–103 jumps in P, 97 pathology (strange asymptotics), xx, 102, 103 posterior error covariance, 101 symmetry of P, 97 kernel, reproducing (rk), 23, 29, 37, 38, 47 kernel, measurement, K , xiv, 46, 140 La Ni˜na, 158 Lagrange multiplier, 14, 25, 41 Lagrangian derivative, 79 Laplacian spline interpolation, xvii likelihood function, 105, 106 linear instability, 74 linear regression, 1, 98 linear system, ﬁnitedimensional, 50 linearization, xvii, 132 dynamical, 74, 76 scheme, experiment with, 80 scheme, possible instability, 74 statistical, 74 local extrema, 15, 16, 19, 24 logarithmic singularity, 52 marine biology, xix Markov chain, 112–114 Markovian correlation, 158 mathematical complexity, xix matrix data error covariance, 40 data weight, 21, 23, 29, 40 deﬁniteness of, 175 diagonal, 48, 49 inverse, 40, 68 kernel, 168 manipulation, 165 notation, 21 sparse, 114 maximum likelihood, xvii, 105, 168 measurement, 12 along a track, 59 general, 45 kernel K , xiv, 46, 140 point, 22, 69 units of, 104 measurement error, 13 covariance, 48, 67, 95 independence of, 64
sample, 73 uncorrelated at different times, 95 variance, 50 variance, stabilizing inﬂuence, 51 vector, 48 measurement functional or operator, xiv, 21, 45, 46, 50, 63, 72, 131, 170 general linear, 48 rotated, 48, 50 measurement site, 24 mechanics, 7 celestial, 125 classical, 8 quantum, 12 meridional overturning circulation, 137 mesoscale motion, 122 mixing front, xviii mixing function, 83, 160 mode, external, 179 mode, internal, 179 model balanced, 144 concept, xiv, 41 convective adjustment, 83 deﬁnition, 44 development, xv dynamically linear, 82 ﬁnitedifference, 136, 166 forward, xiv, 8, 21, 40, 48 grid, actual, 67 grid, coarse, 67 hydrological, 169 improvement, 5 intermediate, 5, 114, 156 layered, 79 shallowwater, 79, 173, 196 mixedlayer, 83 reduced, 144 spectral, 166, 167 statistical nonlinearity, 82 steady, 114 testing, 42, 51 “toy”, xix, 8 unreduced, 146 Modular Ocean Model (MOM), 167, 170 Monte Carlo, estimate, 67 Monte Carlo, Hybrid, 116 Monte Carlo method, xvi, xviii, 64, 74, 160, 169 Montgomery potential, 177, 178 moon, work done by, 138 multiplier, Lagrange, 14, 25, 41 multivariate state, 71 National Aeronautics and Space Administration (NASA) Data Assimilation Ofﬁce, 166 National Center for Atmospheric Research (NCAR), 133
Subject index
National Oceanic and Atmospheric Administration (NOAA), 167 Navy Operational Global Atmospheric Prediction System (NOGAPS), 166 negative eigenvalue, 166, 175 noise, 104, 156 noise, white, 71 noisy inverse, 50 nonsmoothness engineering, 83 ﬁrstorder wave equation, 83 mildest departure from, 84 phase velocity, 83 subgradients, 84 unnatural, 74 normalizing constant, 113 notation, 109, 110 notation, standard, 170 nudge, 149 null subspace, xviii, 50, 51, 74 number Froude, 185 Richardson, 80 Rossby, 123, 142 numerical approximation, 50, 132, 170 grid, coarse, 135 integration, 21 integration, spatial grid for, 99 method, spectral ﬁnite element, 94 modeling, xviii weather prediction (NWP), 156 objective analysis, xviii observable degrees of freedom, 36, 93 observation, 12, 45 observing system, 4, 22, 48, 50 assessment, 5 effectiveness, xv nonlinear, 24 prior error covariances for, 51 ocean bottom, 178 circulation, bestﬁt, xv circulation theory, xviii dynamics, 8 model dynamics, 51 temperature, 105 tide, 124 tropical Paciﬁc, 80, 156–165 open boundary, 127, 180 open loop algorithm, 59, 63 operation count, 65 operator, xiv adjoint, xvii, 167 boundary, 21, 22 expectation, 81
231
generalized inverse, xv initial, 21, 22 inverse, xiv, 21 linear boundary, 130 linear difference, 103 linear differential, 21, 56 linear dynamical, 130 linear model, 56 measurement, 21, 45 nonlinear, 170 nonsingular, xiv, 21 singular, xv, xx, 21 tangentlinear, 167 optimal control theory, xviii, 83 optimal interpolant, 40, 101 optimal interpolation, 37, 38, 40, 54, 104, 111, 169 optimization algorithm, 5, 6 orbit, 128 orthogonality, xviii, 33, 34, 48 outcrop, 79, 183 outﬂow, 175 outputs, 5, 6 Paciﬁc Ocean, tropical, 80, 156–165 parallel processing, 61 paralytic shellﬁsh poisoning, 1 parameter estimation, 168 parameterization, xiii, xv errors, 41 functional nonsmoothness, 74 implausible, 84 particle path, 139 penalty boundary, 18, 22 data, 17, 18 dynamical, 22 initial, 17, 18, 22 roughness, 51–53 penalty function, 115, 166 discontinuous, 116 penalty functional, xv, 14, 16, 17, 19, 21, 25, 28, 33, 34, 46, 50, 52, 53, 75, 131, 134, 139, 163, 168, 180, 182 augmented, 25, 168 evaluation by direct substitution, 89 expected value, 163 in terms of residual, 14, 201 in terms of state variables, 14, 201 minimum value of, 42 nonsmoothness of, 74 N point approximation to, 89 prior value, 24 quartic, 82 reduced or posterior value, 24, 42 perfect dynamics, 90 phase change, xviii phase speed, 8, 75, 141, 160, 168, 179
232
Subject index
physical realizability, 26, 64, 166 point of accumulation, 142 polynomial regression, 5 positive eddy conductivity, 80 positivedeﬁnite matrix, symmetric, 21, 42, 48, 175 posterior error estimate, 3 posterior error statistics, 5 power spectrum, hypothetical, 54 preconditioned conjugate gradient, 146, 166 preconditioner, 63, 64, 72, 78, 92, 151 Primitive Equations, 6, 103, 121, 143, 214 Primitive Equations, reducedgravity, 79 prior covariance of error, 51 data error covariance, 72 data misﬁt, 22, 49, 70, 82 data misﬁt, scaled, 43 rescaling, 163 probability, xix probability, transition, 112, 113 probability distribution function (pdf), 38, 105 conditional, 109 exponential, 106 Gaussian (normal), 114 joint, 109 marginal, 109 nonnormal, 105 normalized, 112 prognostic constraint, xvi Projet d’Assimilation par Logiciel Multim´ethodes (PALM), 171 pseudoheat equation, 65, 167 pseudorandom ﬁeld, 71 input, xvi number generator, 71, 72 sample, 64, 169 quadratic regression, 3 quantum mechanics, 12 quasigeostrophy, 13, 114, 116, 122, 124, 138, 148, 173 radar beam, 45 radiative ﬂux, 46 random error, 37 forcing, 67 input, 42 measurement error, 42 variable, 37, 41, 43, 109 chisquared, 43 Gaussian, 43 observation increases variance, 110 randomization, second, 41, 146 rank, 73
realizability, physical, 26, 64, 166 reciprocalshooting tomography, 45 reference ﬂow, tangent linearization, 77 regression analysis, 6 line, 2 linear, 1, 98 parameters, 2 polynomial, 5 quadratic, 3 regular estimate, 29 reliability, indicator, 67 repeattrack orbit, 134 representer, xvii, 18, 19, 21, 22, 29, 34–36, 53, 131 adjoint, 26, 29, 135 adjoint equation, 19, 69, 204 anisotropy, 153 at different tidal frequencies, 135 coarse grid, 63 crossover difference, 134 family of, xx ﬁne grid, 63 ﬁnite basis, 74 point measurements, 37, 47 residual vector, 70 rotated, 48 vector ﬁeld, 60 representer algorithm, 59, 67, 103 indirect, 61, 63, 72, 78, 135, 146 iterated indirect, 6, 146, 160, 169, 183, 216–220 open loop, 59, 60, 63 representer coefﬁcient, 20, 21, 36, 63, 69, 148 rotated, 49 samples of, 73 representer equation, 20, 69, 72, 204 representer matrix, 21, 24, 48, 60, 63 approximation, 62 eigenvalues, 48–51, 153 eigenvectors, 48–51, 154 ﬁrst K columns, 63, 64 Monte Carlo estimate, 163 m th column of, 61 rank K estimate, 64 spurious asymmetric part, 93 symmetry of, 64, 148 representing property, 34, 35 reprocessed cloud track wind observations (RCTWO), 151 reproducing kernel (rk), 23, 37, 38, 40, 47, 48, 69 residual, xv, 32 bogus, 94 boundary, 130 candidate for, 158 dynamical, 130, 136 minimal, 6 product, 27 weighted, 70 resolution, loss of, 63
Subject index
resolution, poor, 74–75 resolvable ﬁeld, 144 Richardson number, 80 Rhodomonas baltica, 1 Rossby number, 123, 142 sample average, 71 covariance, 64 distribution, 43 estimate, 73 independent, 105 number of, 72, 160 statistics, 183 variance, 106, 108 sample mean, spurious, 72 satellite altimetry, 128 image, 151 orbit, 24, 128 scale analysis, 160 coarse, 145 decorrelation length, 145 decorrelation time, 145 estimate, 145 length, 54, 142 natural, xvii planetary, 167 shallow, 132 synoptic, 143, 148, 167 velocity, 132, 142 sea level, positivity of, 45, 80 Sea Surface Temperature (SST), 5, 156–165 search without gradient information, 85, 115 second randomization, 41, 146 separation of variables, 173, 178 sequential estimation, 94 sigma coordinate, 151, 212 simulated annealing, xviii, 85, 115, 116, 182, 183 singular value decomposition, 5 singularity, 26, 29, 55 skill of forecast, 149, 167 slack constraint, 163 smoothing interval, 98, 103 smoothing norm, 54–56 smoothness, 76 solution continuous dependence, 173 energy, 174 existence, 173 explicit, 21 indeterminate, 215 overdetermined, 174 uniqueness, 9, 84, 172 Southern Ocean, 179
233
space control, xvii, xviii data, 35 null, xviii observable, 36 state, xvii, xviii spline interpolation, Laplacian, xvii SST, see Sea Surface Temperature stability, see conditioning stability, linear, 76 stability criterion, Courant–Friedrich–Lewy, 11 stability of inverse, 49, 50 state variable augmented, 104 multivariate, 71, 130 thermodynamic, 144 virtual, 93 statistical bias, 81 inhomogeneity, 41 interpretation, 42, 66, 168 nonstationarity, 41 prediction scheme, 149 simulation, xvi, 74 stationarity, 102 statistically nonlinear model, 82 Stefan’s law, 46 storage, 61, 73, 74, 98 streamfunction, 13, 45, 123, 138, 142 stress, unresolved, 144 stress, wind, 5 strong constraint, 25, 26, 91, 160, 167 Sturm–Liouville problem, 179 subcritical, 176 suboptimality, 74, 93 supercritical, 175, 176 surface heat ﬂux, 158 stress, 156 temperature, see SST wind, 5, 156–165 survey articles, xx sweep, 115 sweep algorithm, 76, 94 switch, 116, 157 synoptic analysis, xvi tangent linearization, 74, 77, 80 temperature gradient, 80 sea surface, see SST stratospheric, 46 virtual, 94 test independent, 163 model, 6
234
Subject index
test (cont.) signiﬁcance, xv, 3, 43, 74, 163 statistic, 5, 25 theorem central limit, 43 convergence, 142 divergence, 119 embedding, 55 Riesz representation, 33, 46 thermocline, 6, 156–165 thermodynamics, 5, 6, 80, 148, 150, 177 Three Dimensional Variational Assimilation (3DVAR), 166 tide, 124–138 time chart, 22, 60 decorrelation scale, 104 index, 114 line, 149 time scale, 165 evolution, 167 synoptic, 148 time series analysis, xviii time window, 166 timeindependent inverse theory, xx timestepping, 65, 94 tomography, reciprocalshooting, 45 TOPEX/POSEIDON, 124 topology, 155 tracking, radar, 142 tracking, sonar, 142 transition probability, 112, 113 trend, spatial or climatological, 41 trivial solution, 9, 52, 181 Tropical Atmosphere–Ocean (TAO) array, 156–165 Tropical Cyclone 90, experiment, 148 tropical cyclones, xx, 148 Tropical Ocean–Global Atmosphere (TOGA) experiment, 156–165 truncation error, 132, 133 turbulence mixing, 156 problem, 145 weakly homogeneous, 146
unique limit, 142 unique solution, 8, 9, 17, 84, 173–177 United Kingdom Meteorological Ofﬁce, 166 UNIX, xviii, 195 unobservable, 33, 50, 51 uphill search, 116 UTC (Coordinated Universal Time), 152 variance explained, 161 sample, 106 spatial, 65 varianceratio test, 3 variational adjoint method, “THE”, 90 variations, calculus of, xvi, 15, 52, 56, 83, 85, 134 velocity component, 45 irrotational, 144 potential, 143 scale, 134, 142 solenoidal, 144, 148 virtual state variable, 93 vorticity, 114, 123, 138 equation, 123, 124 wave internal, 13, 143 Rossby, 128, 156 topographic, 136 tropical instability, 160 waveguide, equatorial, 158 waverider buoy, 45 wavenumber, 131 weak constraint, 6, 25, 26 weight absolute value, 24 limiting choice, 23 nondiagonal, 29, 53, 60, 181 operator, 51 positive, 52 relative, 24 weighted sum of squared errors (WSSE), 3 wellposed problem, 7, 8, 80, 172 work sheet, 61